refactor(rollout): integrate feedback -> api, recording, log mut and keyboard

feat(rollout): add subtask command
feat(rollout): mute logs in interactive mode
2026-08-08 17:39:44 +00:00 · 2026-08-07 21:33:17 +02:00 · 2026-08-07 16:44:29 +02:00 · 2026-08-07 16:00:56 +02:00 · 2026-08-07 15:22:44 +02:00 · 2026-08-07 13:52:23 +02:00
165 changed files with 12201 additions and 3867 deletions
@@ -53,7 +53,7 @@ permissions:
  contents: read

 env:
-  UV_VERSION: "0.8.0"
+  UV_VERSION: "0.11.30"
  PYTHON_VERSION: "3.12"

 # Cancel in-flight runs for the same branch/PR.
@@ -27,7 +27,7 @@ on:

 # Sets up the environment variables
 env:
-  UV_VERSION: "0.8.0"
+  UV_VERSION: "0.11.30"
  PYTHON_VERSION: "3.12"
  DOCKER_IMAGE_NAME_CPU: huggingface/lerobot-cpu:latest
  DOCKER_IMAGE_NAME_GPU: huggingface/lerobot-gpu:latest
@@ -48,7 +48,7 @@ permissions:

 # Sets up the environment variables
 env:
-  UV_VERSION: "0.8.0"
+  UV_VERSION: "0.11.30"
  PYTHON_VERSION: "3.12"

 # Ensures that only the latest commit for a PR or branch is built, canceling older runs.
@@ -37,7 +37,7 @@ permissions:

 # Sets up the environment variables
 env:
-  UV_VERSION: "0.8.0"
+  UV_VERSION: "0.11.30"
  PYTHON_VERSION: "3.12"
  DOCKER_IMAGE_NAME: huggingface/lerobot-gpu

@@ -27,7 +27,7 @@ on:

 # Sets up the environment variables
 env:
-  UV_VERSION: "0.8.0"
+  UV_VERSION: "0.11.30"
  PYTHON_VERSION: "3.12"
  DOCKER_IMAGE_NAME: huggingface/lerobot-gpu:latest-deps

@@ -21,7 +21,7 @@ on:

 # Sets up the environment variables
 env:
-  UV_VERSION: "0.8.0"
+  UV_VERSION: "0.11.30"
  PYTHON_VERSION: "3.12"

 jobs:
@@ -19,8 +19,8 @@ on:
  workflow_dispatch:

  # Runs at 02:00
-  # schedule:
-  #   - cron: "0 2 * * *"
+  schedule:
+    - cron: "0 2 * * *"

 env:
  CLOSE_ISSUE_MESSAGE: >
@@ -31,7 +31,7 @@ env:
    Feel free to reopen if is still relevant, or to ping a collaborator if you have any questions.
  WARN_ISSUE_MESSAGE: >
    This issue has been automatically marked as stale because it has not had
-    recent activity (1 year). It will be closed if no further activity occurs.
+    recent activity (1 year). It will be closed if no further activity occurs within 30 days.
    Any change, comment or update to this issue will reset this count.
    Thank you for your contributions.
  WARN_PR_MESSAGE: >
@@ -61,8 +61,8 @@ jobs:
          exempt-pr-labels: never-stale
          days-before-issue-stale: 365
          days-before-issue-close: 30
-          days-before-pr-stale: 365
-          days-before-pr-close: 30
+          days-before-pr-stale: -1
+          days-before-pr-close: -1
          delete-branch: true
          close-issue-message: ${{ env.CLOSE_ISSUE_MESSAGE }}
          close-pr-message: ${{ env.CLOSE_PR_MESSAGE }}
@@ -0,0 +1,457 @@
+# Interactive Rollout — Design Notes
+
+Branch: `feat/add_interactive_rollout` · Status: Phases 1–2 committed; Round 2
+(programmatic API, sentry support, muting v2, stdin move) implemented and tested,
+uncommitted.
+
+---
+
+## 1. Vision
+
+`lerobot-rollout` runs inference on a real robot: it connects hardware, loads the policy,
+builds the processor pipelines, optionally records a dataset, and spins the control loop.
+Today that is a **one-shot, fire-and-forget** program. You pass `--task="pick up the cube"`
+on the command line, the robot starts moving immediately, and the only interaction left is
+Ctrl-C. If you want a different instruction, you kill the process and pay the full startup
+cost again — reconnecting motors, re-homing, re-loading a multi-GB VLA onto the GPU.
+
+Since LeRobot gained subtask annotation and language conditioning, that model is the
+bottleneck. The **north star** is a chat-style CLI over stdin, where the operator issues
+commands *concurrently with the robot moving*:
+
+```
+/start                              begin (or resume) the policy control loop
+/subtask Grab the red cube          re-instruct the policy on the fly
+/ask what's the capital of France?  query an LLM while the robot keeps moving
+/reset                              stop movement, return home, clear the subtask —
+                                    but keep hardware and policy warm
+/stop                               graceful shutdown
+```
+
+The unifying idea: **the expensive things (hardware, policy weights, processors) stay warm
+across commands.** Only the cheap things — the instruction, the control loop — start and
+stop. That turns a rollout from a batch job into a session you can steer.
+
+## 2. Objective (scoped)
+
+Phased, so each phase lands as a reviewable unit:
+
+| Phase | Scope | Status |
+|---|---|---|
+| **1** | `--interactive` flag, non-blocking stdin listener, command parser, `/start` `/reset` `/stop` `/help` | ✅ done |
+| **1.5** | Mute system logs so they stop fighting the prompt for the terminal | ✅ done |
+| **2** | `/subtask <text>` — change the policy's instruction mid-run | ✅ done |
+| **2.5** | Round 2: `RolloutController` public API, sentry recording support, muting v2 (errors surface), stdin listener → `lerobot/utils` | ✅ done (see §5) |
+| **3** | `/ask` + hierarchical task-vs-subtask semantics (LLM in the loop) | not started |
+
+An explicit constraint through Phases 1–2: **do not couple this to the language runtime yet.**
+Build the mechanism; keep the door open.
+
+## 3. Inspiration — three reference PRs
+
+We read all three and deliberately implemented none of them verbatim.
+
+**PR #4108 — online subtask switching.** Introduces a `PromptBroker` + `PromptListenerBase`
+ `StdinPromptListener`, a `RuntimeContext.prompt_broker` field, `register_on_change`
+callbacks, an `--online_task_switching_flush` config flag, and `flush_action_queue()` /
+`_apply_pending_flush()` on `PreTrainedPolicy` — **with edits to 14 policy files** to call
+the flush at the top of `select_action`. Its architecture is designed for pluggable input
+sources (network, voice), which is the right long-term shape but more machinery than we
+need. *What we took:* the core insight that a mid-run instruction change must invalidate
+actions precomputed under the old instruction, and that the flush must happen on a thread
+that is safe to touch policy state from.
+
+**PR #4183 — experimental full-UX draft.** Achieves the whole north-star vision, but does
+so by adding a `lerobot.runtime` / `language_runtime.py` that **duplicates** `BaseStrategy`,
+`send_next_action`, and the rollout control loop. *What we took:* the UX target and the
+command vocabulary. *What we rejected:* the parallel runtime — a second control loop is a
+second thing to keep correct, and everything it does is already in `rollout/strategies/`.
+
+**PR #4234 — policy-side edits enabling #4183's runtime.** Read for context on where the
+language plumbing lands inside a policy. Relevant to Phase 3, not to what we built.
+
+## 4. What we built, and why
+
+Three commits on the branch:
+
+```
+072c697c0  feat(rollout): interactive v1
+d3ee0b820  feat(rollout): mute logs in interactive mode
+39c4e746f  feat(rollout): add subtask command
+```
+
+Cumulative footprint — one new module, one new test file, small surgical edits elsewhere:
+
+```
+ src/lerobot/rollout/interactive.py         | 580 +++++   (new)
+ tests/test_interactive_rollout.py          | 788 +++++   (new)
+ docs/source/inference.mdx                  |  87 +++
+ src/lerobot/rollout/inference/base.py      |  66 +++
+ src/lerobot/rollout/inference/rtc.py       |  61 +-
+ src/lerobot/rollout/inference/sync.py      |  21 +-
+ src/lerobot/scripts/lerobot_rollout.py     |  32 +-
+ src/lerobot/policies/pretrained.py         |  24 +
+ src/lerobot/rollout/strategies/core.py     |  21 +-
+ src/lerobot/rollout/configs.py             |  18 +
+ src/lerobot/rollout/__init__.py            |  16 +-
+ src/lerobot/rollout/strategies/episodic.py |   4 +-
+```
+
+The ratio matters: **~1400 of ~1680 added lines are the new module and its tests.** The
+existing rollout architecture was reused, not reshaped.
+
+### 4.1 Segments over a linked event — the load-bearing idea
+
+Every rollout strategy's control loop already polls `ctx.runtime.shutdown_event.is_set()`
+to know when to stop. So instead of teaching strategies about interactivity, we **swap in a
+smarter event**:
+
+```python
+class LinkedEvent(Event):
+    """is_set() reflects the local flag OR a parent event."""
+    def is_set(self) -> bool:
+        return super().is_set() or self.parent.is_set()
+```
+
+`lerobot-rollout` wraps the `ProcessSignalHandler`'s shutdown event in a `LinkedEvent` when
+`--interactive=true`. The session sets the **local** flag to end a run *segment*; SIGINT /
+SIGTERM still arrive through the **parent**, so Ctrl-C behaves exactly as before.
+
+`InteractiveSession.run()` then drives `strategy.run(ctx)` in restartable segments:
+
+```
+setup(ctx)  →  [idle]  →  /start → run(ctx) → /reset → [idle] → /start → run(ctx) → /stop  →  teardown(ctx)
+                                                  ↑ hardware + policy stay warm throughout
+```
+
+**Zero strategy code changed** to support this. The only additions to `strategies/core.py`
+were `reset_control_state()` (engine + interpolator + cached-observation reset, factored
+out of `_init_engine` so a segment can restart cleanly) and making
+`_return_to_initial_position` public.
+
+### 4.2 Threading model
+
+```
+ listener thread  ──publishes flags / strings──▶  main thread
+ (stdin reader)      never touches hardware        (session loop → strategy.run → control loop)
+                     never mutates policy state
+```
+
+The listener only ever writes `threading.Event` flags and a lock-guarded string. Everything
+that touches hardware or policy state happens on the thread that already owns it. This
+mirrors the existing DAgger events pattern rather than inventing a new concurrency idiom.
+
+### 4.3 stdin must be read with `os.read`, not `readline`
+
+Non-obvious and load-bearing. The first implementation used `select()` + `stream.readline()`
+and **two tests failed**: a buffered file object slurps *several* lines off the file
+descriptor in one syscall, after which `select` reports the drained fd as not-ready and the
+buffered lines are never delivered. Pasted or piped command batches got stuck. The reader
+now does `select()` + `os.read(fd, 4096)` + manual `\n` splitting, with a
+blocking-`readline` fallback for streams without a `fileno()` (non-POSIX, test doubles).
+
+Also: unlike `TerminalKeyListener`, this reader leaves the terminal in **canonical mode** —
+the operator is typing chat commands, not pressing hotkeys.
+
+### 4.4 EOF means stop
+
+A closed stdin means there is no way left to command the robot, so EOF (Ctrl-D, or an
+exhausted piped script) stops the session. An unexpected read error is treated the same way,
+for the same reason. Consequence, documented: piped scripts must hold stdin open —
+
+```bash
+(printf '/start\n'; sleep 60; printf '/stop\n') | lerobot-rollout ... --interactive=true
+```
+
+### 4.5 Commands are last-write-wins
+
+`/reset` and `/stop` cancel a still-pending `/start`, so the robot never starts moving after
+the operator's most recent command said not to. Handlers set their intent flag *first* and
+the segment-stop event *second*; `_run_segment` clears the segment-stop flag *before*
+re-checking the intent flags. A `/reset` racing a `/start` is therefore either seen before
+the segment begins or ends it on its first tick.
+
+### 4.6 Base strategy only (enforced by config validation)
+
+`--interactive=true` with a recording strategy raises a `ValueError`. Two reasons: recording
+strategies finalize their dataset inside `run()` (so `run()` is not restartable), and their
+keyboard listeners contend with the command reader for the same TTY. This is a deliberate,
+documented limitation — not an oversight.
+
+### 4.7 Log muting (Phase 1.5)
+
+Policy, robot and control-loop logs at every level interleave with the chat prompt and
+destroy the typing UX. Simplest workable answer, per explicit request: **mute console output
+for the duration of the session.**
+
+- Every logger's console `StreamHandler` is raised above `CRITICAL` — **not just root**,
+  because `transformers` and `datasets` attach their own stderr handlers with
+  `propagate=False`.
+- `warnings.simplefilter("ignore")`, with `warnings.filters` saved and restored.
+- **File handlers are untouched** — anyone wanting a persistent log can attach one.
+- Restored in `run()`'s `finally`, *before* the closing `log_say`, so teardown logs are visible.
+
+The obvious hazard: muting hides fatal errors. So `InferenceEngine` gained a
+`failure_traceback` property, RTC captures its traceback in the fatal handler, and the
+session prints it on failure. **Do not remove that when touching the failure path.**
+
+"See both logs and prompt" — a pinned input line, `prompt_toolkit`-style — was deliberately
+deferred: it needs a new dependency and a real TUI layer.
+
+### 4.8 `/subtask` — the engine *is* the broker
+
+The pivotal call on Phase 2: **skip PR #4108's `PromptBroker`.** After Phase 1, the session
+already owns the stdin thread and the parser, so a broker + listener base + on-change
+callbacks + a new `RuntimeContext` field would be duplicate machinery — and callbacks firing
+on the listener thread are exactly the cross-thread hazard we designed against.
+
+Instead, `InferenceEngine` (the ABC every backend already implements) became the thread-safe
+task holder:
+
+```python
+@property
+def task(self) -> str: ...              # lock-guarded read
+
+def set_task(self, task) -> bool:       # callable from ANY thread; True if it changed
+    ...
+
+def _take_task(self) -> tuple[str, bool]:   # consumed on the INFERENCE thread;
+    ...                                     # returns (task, changed) and clears the edge
+```
+
+`/subtask` is then three lines: read `engine.task`, call `engine.set_task(text)`, print the
+transition. No new module, no new context field, no callbacks.
+
+**The flush problem, and why it got small.** When the instruction changes, a chunking policy
+is still serving actions computed under the old one — up to `chunk_size` ticks of stale
+behavior. PR #4108 solved this by adding `flush_action_queue()` / `_apply_pending_flush()`
+to `PreTrainedPolicy` **and editing 14 policy files**, because its flush request arrived from
+a foreign thread and had to be deferred to a safe point inside `select_action`.
+
+Ours already runs *on* the thread that calls `select_action`. So: one concrete method on
+`PreTrainedPolicy` and **zero per-policy edits**.
+
+```python
+def drop_queued_actions(self) -> None:
+    queues = getattr(self, "_queues", None)
+    if isinstance(queues, dict) and ACTION in queues:
+        queues[ACTION].clear()
+    action_queue = getattr(self, "_action_queue", None)
+    if action_queue is not None:
+        action_queue.clear()
+```
+
+Two `getattr`s cover the repo's two queue idioms across all ~18 policies
+(`_queues[ACTION]`: diffusion, smolvla, tdmpc, vqbet, wall_x, xvla, multi_task_dit, vla_jepa;
+`_action_queue`: act, pi0, pi05, pi0_fast, eo1, evo1, groot, molmoact2, fastwam, lingbot_va).
+Policies with no queue inherit a no-op.
+
+**Why not `policy.reset()`?** That was the first implementation, and review caught it as too
+blunt. For Diffusion it wipes the observation history, so the next chunk is planned from a
+history of the current frame repeated — a visible discontinuity mid-motion. And ACT /
+Diffusion / VQBeT / TDMPC don't read `task` at all, so they'd pay that jerk for nothing.
+`drop_queued_actions` keeps episode state and drops only what is actually stale.
+
+**RTC deliberately does *not* flush.** Clearing its queue would leave the robot with no
+commands for a full inference latency (~1 s on a VLA). Instead the next chunk is generated
+under the new instruction and merged over the previous chunk's leftover prefix — the switch
+lands within one inference and the motion stays continuous. That is exactly what RTC's
+blending exists for. Documented per-backend in `inference.mdx`; no config flag, one sensible
+default per backend.
+
+**`/reset` restores the launch task on the listener thread.** Subtle and worth preserving:
+the restore lives in `_cmd_reset`, not in `_reset_robot` (which runs later, on the main
+thread). Otherwise `/reset` followed immediately by `/subtask` would be ordered by *service*
+time rather than *command* time, and the deferred restore would silently revert the new
+instruction — deterministically so, for pasted or piped input. Both writers now run on the
+same thread, so command order wins. There is a regression test driving this through a real pipe.
+
+## 5. Round 2 — the feature becomes a library API
+
+Four follow-up asks landed together (currently uncommitted on the branch):
+make the components programmatic-API friendly (the priority), extend interactive
+to recording where cheap, simplify muting / surface errors, and settle the
+ssh/headless + `keyboard_input` question.
+
+### 5.1 `RolloutController` — programmatic control
+
+`interactive.py` bisected cleanly, so the generic control logic moved to a new
+`rollout/controller.py`:
+
+```python
+controller = RolloutController(strategy, ctx, on_event=my_observer)
+controller.serve()      # blocking loop (run it on whatever thread you like)
+controller.start()      # -> bool: False when a segment is already running
+controller.set_task(t)  # -> bool: re-instruct mid-run, from any thread
+controller.reset()      # -> bool: True when the launch task was restored
+controller.stop()
+controller.task / .initial_task / .running / .failed / .failure_traceback
+```
+
+- **No I/O of its own** — no stdin, no prints, no log muting, no TTS. Every
+  state transition that used to be a `print` is now a `RolloutEvent`
+  (`SEGMENT_STARTED`, `SEGMENT_ENDED`, `RESET_STARTED/DONE/SKIPPED`,
+  `ENGINE_FAILED`, `STOPPED`) emitted on the serve thread.
+- **Thread-safe by lock, not by convention.** The old ordering guarantee
+  (`/subtask` right after `/reset` must win) relied on both writes running on
+  the single stdin thread. The controller serializes `start`/`reset`/`stop`/
+  `set_task` with an internal lock, so the guarantee now holds for arbitrary
+  caller threads — the prerequisite for network/voice front-ends.
+- `InteractiveSession` shrank to a thin adapter: stdin listener + parser +
+  rendering + muting; each command maps 1:1 onto a controller method, and the
+  controller is exposed as `session.controller`.
+- Exported from `lerobot.rollout`: `RolloutController`, `RolloutEvent`,
+  `LinkedEvent`. `docs/source/inference.mdx` gained a **Programmatic control**
+  section with a complete embedding example.
+
+### 5.2 Sentry + interactive — recording while you steer
+
+Decision, per the agreed criteria: the `/record` keyboard-handoff idea is
+**medium-to-large** (listeners have no suspend/resume API and start at
+creation, `esc` handlers are hardcoded and collide, pynput captures globally
+while you type, and each strategy carries per-run stale flags) → rejected.
+But the investigation showed **sentry has zero keyboard code** — the config
+comment lumping it with the keyboard strategies was simply wrong — and its
+only real blocker was one line: `with VideoEncodingManager(dataset)` inside
+`run()` finalizes the dataset the first time `run()` returns, after which a
+restarted segment would silently truncate the finalized parquet.
+
+So `--interactive=true` now supports `--strategy.type=sentry`:
+
+- **Finalization moved to `teardown()`** (which already called
+  `dataset.finalize()`); `run()` is segment-restartable. Each segment saves
+  complete episodes plus one tail partial episode; on a failed tail save the
+  in-flight streaming encode is cancelled *and* the half-mutated episode
+  buffer is discarded (see §6, round 2).
+- **Frames are labeled with the live `engine.task`** instead of a config
+  snapshot — the writer already stores a task per frame — so `/subtask`
+  changes the policy conditioning and the recorded label from the same frame
+  onward. This also resolved the "recorded frames ignore `/subtask`" open item
+  for sentry.
+- `episodes_since_push` hoisted to instance state so upload cadence survives
+  segments.
+- dagger / highlight / episodic stay excluded: keyboard conflicts plus per-run
+  recording state that does not survive a restart.
+
+### 5.3 Muting v2 — two lines, and errors surface
+
+The ~30-line per-handler walk became `logging.disable(logging.WARNING)` with
+the previous disable level restored afterwards. Strictly better coverage: the
+gate applies before handler dispatch, so it covers `propagate=False` library
+loggers *and* loggers created mid-session (the old snapshot missed those) —
+and **ERROR/CRITICAL now reach the console**, which the audit showed is safe:
+no ERROR-level emitter fires periodically in healthy operation (the periodic
+nuisances — slow-loop, camera hiccups — are WARNINGs and stay muted).
+Documented trade-off: the gate also withholds INFO/WARNING from file handlers
+during the session; acceptable because no default code path attaches one
+(only `rl/actor`, `rl/learner`, `async_inference` pass `log_file`). The
+`warnings` suppression stays (nothing calls `logging.captureWarnings`), and
+`failure_traceback` surfacing stays as the belt-and-suspenders for fatal
+engine errors.
+
+### 5.4 stdin listener → `lerobot/utils/stdin_input.py`
+
+The ssh/headless audit confirmed the listener was already the right design:
+`select`+`os.read` works over SSH (the session pty is a normal fd), from
+pipes, and headless — it's `keyboard_input`'s **pynput** backend that needs a
+display server. Nothing in `keyboard_input` overlaps enough to reuse
+(1-byte cbreak hotkey decoding vs canonical-mode line assembly), so
+`StdinCommandListener` moved to a **new** utils module — deliberately not
+into `keyboard_input.py`, which attempts a pynput import at module load.
+Canonical import only: `lerobot.utils.stdin_input` (removed from
+`lerobot.rollout`'s exports).
+
+The move fixed a real bug the audit found: with `sys.stdin is None`
+(daemonized processes), the blocking fallback died with an uncaught
+`AttributeError` without firing `on_eof` — leaving a session idling with no
+command channel. `start()` now treats a missing stream as immediate EOF.
+
+## 6. Bugs the adversarial reviews caught
+
+Four multi-agent review passes were run across the phases (28 / 5 / 27 / 12 agents;
+findings adversarially verified before acting). The ones that mattered:
+
+**Round 2 (2 confirmed, 0 refuted):**
+
+- **Controller `start()` race → phantom segment.** `start()` gated on `_running`, but the
+  serve loop cleared `_start_requested` *before* setting `_running` — a second `start()`
+  landing in that window (spanning `reset_control_state` and the SEGMENT_STARTED emission)
+  returned `True` and re-armed the flag, which nothing consumed during the segment; the
+  robot would start again, uncommanded, when the segment later ended on its own. Fixed:
+  the serve loop consumes the request and sets `_running` atomically under the control
+  lock, and `_running` spans the whole startup sequence.
+- **Sentry poisoned episode buffer.** `save_episode` mutates the buffer in place (pops
+  `size`/`task`) *before* the fallible writes; a failed tail save left a half-mutated dict
+  and the next segment's first `add_frame` crashed with `KeyError('size')`. Fixed: the
+  except branch discards the buffer so `add_frame` recreates it.
+
+**Rounds 1–3 (Phases 1–2):**
+
+- **RTC stale observation (critical).** `RTCInferenceEngine.reset()` never cleared
+  `_obs_holder["obs"]`. After `/reset` physically moved the robot home, the next `/start`
+  computed its first chunk from the **pre-reset pose** — a lurch back toward where the arm
+  used to be. Fixed by clearing the observation and adding a `_reset_epoch` counter so an
+  in-flight chunk computed across a reset is discarded rather than merged. This also fixes a
+  pre-existing DAgger staleness path.
+- **Muting hid fatal errors** → `failure_traceback` capture + session print (§4.7).
+- **Muting scope too narrow** → root-only missed `transformers` / `datasets`; `warnings`
+  output bypassed logging entirely.
+- **Command ordering** → `/reset` and `/stop` didn't cancel a pending `/start` (§4.5); the
+  `/reset`-then-`/subtask` clobber (§4.8).
+- **Flush too heavy** → `policy.reset()` → `drop_queued_actions()` (§4.8).
+- **Empty-task rendering** → `''` replaced with `(none — set one with /subtask <text>)`.
+- **Silent switch** → the confirmation now says "(applies from the next policy inference)",
+  since the explanatory logs are muted.
+
+## 7. Verification
+
+After Round 2:
+
+```
+uv run --extra dataset pytest tests/test_interactive_rollout.py \
+    tests/utils/test_stdin_input.py tests/test_rollout.py -q
+    → 81 passed
+
+pre-commit (all changed files)
+    → 0 failures
+```
+
+Phase 1–2 numbers (still green at the time): 64 rollout/interactive tests;
+223 passed / 5 skipped across `tests/policies/rtc`, factory, and common
+(confirming the shared `pretrained.py` change); pre-commit 0 failures.
+
+`tests/test_interactive_rollout.py` covers the parser, `LinkedEvent` semantics,
+`RolloutController` (start/reset/stop/set_task flows, events, startup-race
+rejection, failure surfacing, broken observers), session flows (start / reset /
+restart / stop, cancel-pending-start, engine failure with traceback, natural
+end, EOF, and a real `BaseStrategy` end-to-end), muting (INFO/WARNING blocked,
+ERROR surfaces, pre-existing disable level restored), `/subtask` semantics,
+sentry restartability + live labels + failed-tail-save recovery, the engine
+task holder, the sync flush, and `drop_queued_actions`.
+`tests/utils/test_stdin_input.py` covers the listener (select path, batched
+lines, blocking fallback, EOF, handler errors, None-stdin, broken streams).
+
+## 8. Extension points for Phase 3
+
+The design was built to make `/ask` an additive change:
+
+- **Command table.** `InteractiveSession._commands` is `name → (handler, arg hint, help)`.
+  `/help` and the startup banner render from it, so a new command is documented for free.
+- **Controller API.** New front-ends (network, voice, `/ask`'s LLM worker) call
+  `RolloutController.start/reset/stop/set_task` from their own threads — the internal lock
+  makes that safe — and observe `RolloutEvent`s instead of scraping terminal output.
+- **Thread discipline.** A command handler runs on the listener thread and must only call
+  controller methods. An LLM call belongs on its own worker thread so the robot keeps
+  moving — precisely the concurrency `/ask` is meant to demonstrate.
+- **Task holder.** `set_task` / `_take_task` already give any producer a safe way to
+  re-instruct the policy. Hierarchical task-vs-subtask semantics (per #4183 / #4234) layer
+  on top of it rather than replacing it.
+
+Open items, deliberately not addressed:
+
+- dagger / highlight / episodic remain non-interactive (keyboard conflicts + per-run
+  recording state); they also still snapshot the task label per run. Sentry is the
+  supported recording path for interactive sessions.
+- The "see logs and prompt simultaneously" TUI (pinned input line).
+- Non-stdin input sources (network, voice) — now unblocked by `RolloutController`; #4108's
+  pluggable-listener shape remains the reference for the transport layer.
@@ -128,6 +128,23 @@ lerobot-eval \

 Learn how to implement your own simulation environment or benchmark and distribute it from the HF Hub by following the [EnvHub Documentation](https://huggingface.co/docs/lerobot/envhub).

+### Third-Party Hardware
+
+Beyond the natively supported hardware, the community maintains a growing ecosystem of plugins for other robots, teleoperators, cameras, and sensors - UFACTORY xArm, Universal Robots UR5e, Franka, AgileX Piper, Trossen WidowX, ARX5, I2RT YAM, GELLO, SpaceMouse, Meta Quest, ROS 2 bridges, tactile and depth cameras, and more.
+
+Plugins are auto-discovered by package name: LeRobot imports any installed package prefixed with `lerobot_robot_`, `lerobot_teleoperator_`, or `lerobot_camera_`. Install one and use the `type` it registers straight from the CLI:
+
+```bash
+pip install lerobot_robot_<name> lerobot_teleoperator_<name>
+
+lerobot-record \
+  --robot.type=<robot_name> \
+  --teleop.type=<teleoperator_name> \
+  --dataset.repo_id=${HF_USER}/my-dataset
+```
+
+Browse the full list in the [Third-Party Robots & Teleoperators](https://huggingface.co/docs/lerobot/main/third_party_robots) and [Third-Party Cameras & Sensors](https://huggingface.co/docs/lerobot/main/third_party_sensors) documentation.
+
 ## Resources

 - **[Documentation](https://huggingface.co/docs/lerobot/index):** The complete guide to tutorials & API.
@@ -165,6 +165,8 @@
    title: OpenArm
  - local: rebot_b601
    title: reBot B601-DM
+  - local: third_party_robots
+    title: Third-Party Robots & Teleoperators
  title: "Robots"
 - sections:
  - local: phone_teleop
@@ -175,6 +177,8 @@
 - sections:
  - local: cameras
    title: Cameras
+  - local: third_party_sensors
+    title: Third-Party Cameras & Sensors
  title: "Sensors"
 - sections:
  - local: notebooks
@@ -161,6 +161,16 @@ The methods called by the train/eval loops:

 Batches are flat dictionaries keyed by the constants in [`lerobot.utils.constants`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/utils/constants.py): `OBS_STATE` (`observation.state.<motor>`), `OBS_IMAGES` (`observation.images.<camera>`), `OBS_LANGUAGE`, `ACTION`, etc. Reuse the constants — don't invent new prefixes.

+If your model is large enough to warrant [sharded multi-GPU training](./multi_gpu_training#sharded-training-fsdp), also declare its FSDP wrap units — the repeated block classes sharding operates on:
+
+```python
+class MyPolicy(PreTrainedPolicy):
+    ...
+    _fsdp_wrap_modules = ["MyTransformerBlock"]
+```
+
+With this one declaration, `--parallelism.dp_shard=N` works out of the box for your policy (users can still override it with `--accelerator.fsdp.wrap_modules`). Without any wrap source, sharded runs fail at startup by design.
+
 ### Processor functions

 LeRobot uses `PolicyProcessorPipeline`s to normalize inputs and de-normalize outputs around your policy. For a concrete reference, see [`processor_act.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/act/processor_act.py) or [`processor_diffusion.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/diffusion/processor_diffusion.py).
@@ -300,7 +310,7 @@ The file names are load-bearing: the factory does lazy imports by name, and the
 Two places need to know about your policy. All by name.

 1. **`policies/__init__.py`** — re-export `MyPolicyConfig` and add it to `__all__`. This import is what registers your policy: `@PreTrainedConfig.register_subclass("my_policy")` runs, and from then on the factory resolves everything by convention. **Don't** re-export the modeling class; it loads lazily through the factory (so `import lerobot` stays fast).
-2. **`templates/lerobot_modelcard_template.md` and the root `README.md`** — the template is what `push_model_to_hub` renders into the model card of every checkpoint trained with your policy: add a one-line description of your policy in the `model_name` branches, map it in `policy_docs` so cards link to your MDX guide, and optionally add an architecture image to `diagrams`. Then add your policy to the models table in the root `README.md`, under the right category, linking to your doc page.
+2. **`templates/lerobot_modelcard_template.md` and the root `README.md`** — the template is what the end-of-training publisher renders into the model card of every checkpoint trained with your policy: add a one-line description of your policy in the `model_name` branches, map it in `policy_docs` so cards link to your MDX guide, and optionally add an architecture image to `diagrams`. Then add your policy to the models table in the root `README.md`, under the right category, linking to your doc page.

 Mirror an existing policy that's structurally similar to yours; the diff is small.

@@ -344,7 +354,7 @@ A new policy is much easier to review — and far more useful — when it ships

 **Pick at least one in-tree benchmark.** LeRobot ships sim benchmarks with per-benchmark Docker images (LIBERO, LIBERO-plus, Meta-World, RoboTwin 2.0, RoboCasa365, RoboCerebra, RoboMME, VLABench and more). Pick the one that matches your policy's modality — VLAs usually go to LIBERO or VLABench; image-only BC to LIBERO or Meta-World. The full list lives under [Benchmarks](./libero) in the docs sidebar.

-**Push the checkpoint & processors** to the Hub under `lerobot/<policy>_<benchmark>` (or your namespace if you don't have write access; a maintainer can mirror it). Use `PreTrainedPolicy.push_model_to_hub` so the repo gets `config.json`, `model.safetensors`, and a model card.
+**Push the checkpoint & processors** to the Hub under `lerobot/<policy>_<benchmark>` (or your namespace if you don't have write access; a maintainer can mirror it). The easiest way is training with `--policy.repo_id=<namespace>/<repo>` and `--policy.push_to_hub=true`: `lerobot-train` publishes the model, both processors, and a model card at the end of the run. To publish an existing checkpoint after the fact, upload its `pretrained_model/` directory (e.g. `huggingface-cli upload`), or use `lerobot-convert-dcp --push_to_hub=...` for sharded-format checkpoints.

 **Report results in your policy's MDX**, with the exact `lerobot-eval` command and hardware so anyone can re-run:

@@ -23,18 +23,18 @@ The broader EVO1 project may include additional training scripts and dataset too
 2. Install EVO1 dependencies:

   ```bash
-   pip install -e ".[evo1]"
+   pip install -e ".[training,evo1]"
   ```

-   For LIBERO evaluation, install the LIBERO extra as well:
+   For LIBERO training and evaluation, install the LIBERO extra as well:

   ```bash
-   pip install -e ".[evo1,libero]"
+   pip install -e ".[training,evo1,libero]"
   ```

 3. Install a `flash-attn` wheel only if it is compatible with your Python, PyTorch, CUDA, and GPU stack. EVO1 falls back to standard attention when `flash_attn` is not available.

-EVO1 uses the native Hugging Face `transformers` InternVL implementation, so `policy.vlm_model_name` must point to a natively converted checkpoint such as `OpenGVLab/InternVL3-1B-hf` (note the `-hf` suffix). The first run may download the configured VLM checkpoint unless `policy.vlm_model_name` points to a local model directory.
+EVO1 uses the native Hugging Face `transformers` InternVL implementation, so `policy.vlm_model_name` must point to a natively converted checkpoint such as `OpenGVLab/InternVL3-1B-hf` (note the `-hf` suffix). The first run downloads the configured VLM checkpoint and later runs reuse it from the Hugging Face cache.

 ## Data Requirements

@@ -92,7 +92,7 @@ lerobot-train \

 ### Stage 2

-Stage 2 finetunes the VLM branches and action head. A common workflow starts from a Stage 1 checkpoint:
+Stage 2 loads the Stage 1 policy, but starts a fresh optimizer and scheduler:

 ```bash
 lerobot-train \
@@ -152,16 +152,154 @@ lerobot-rollout \

 ### LIBERO Evaluation

-> [!NOTE]
-> Benchmark results for a `lerobot`-hosted LIBERO checkpoint trained with this implementation
-> will be added once training completes.
+#### Reference result

-The official EVO1 LIBERO rollout protocol uses the raw LIBERO camera feature names
+> [!NOTE]
+> The released Stage-2 checkpoint passed clean-download and rollout verification:
+> [`zuoxingdong/evo1_libero`](https://huggingface.co/zuoxingdong/evo1_libero), revision
+> [`515921f4a2c1d3f3ad523721eafa26fdf2af315b`](https://huggingface.co/zuoxingdong/evo1_libero/commit/515921f4a2c1d3f3ad523721eafa26fdf2af315b).
+> The clean-download evaluation used LeRobot revision
+> [`e40b58a8dfa9e7b86918c374791599d070518d11`](https://github.com/huggingface/lerobot/commit/e40b58a8dfa9e7b86918c374791599d070518d11).
+
+The single-run Stage-2 checkpoint at step 70,000 produced:
+
+| Suite          | Successful episodes |  Episodes | Success rate |
+| -------------- | ------------------: | --------: | -----------: |
+| LIBERO Spatial |                 485 |       500 |        97.0% |
+| LIBERO Object  |                 496 |       500 |        99.2% |
+| LIBERO Goal    |                 483 |       500 |        96.6% |
+| LIBERO-10      |                 469 |       500 |        93.8% |
+| **Overall**    |           **1,933** | **2,000** |   **96.65%** |
+
+These results use one trained checkpoint and evaluation seed `1000`; they are not a multi-seed
+mean or confidence estimate.
+
+#### Reference training recipe
+
+The released checkpoint records the complete resolved Stage-2 configuration in
+[`train_config.json`](https://huggingface.co/zuoxingdong/evo1_libero/blob/515921f4a2c1d3f3ad523721eafa26fdf2af315b/train_config.json).
+The measured run used two H100 GPUs with two DDP processes and batch 64 per process, giving global batch 128. Both stages used the same topology. The base VLM came from revision
+`014c0583a0d4bedf29fbe2dbff4f865eb998e171` of `OpenGVLab/InternVL3-1B-hf`.
+The released artifact does not record the exact LeRobot training commit or its original dependency lock,
+so the commands below reproduce the recorded configuration and topology from a current checkout rather
+than reconstructing the software environment bit for bit.
+
+From a LeRobot source checkout, install the locked dependencies and download that exact VLM revision:
+
+```bash
+uv sync --locked --extra training --extra evo1 --extra libero
+VLM_DIR=$(uv run hf download OpenGVLab/InternVL3-1B-hf \
+  --revision=014c0583a0d4bedf29fbe2dbff4f865eb998e171)
+```
+
+Stage 1 freezes the VLM and trains the action head for 5,000 steps:
+
+```bash
+uv run accelerate launch --num_processes=2 -m lerobot.scripts.lerobot_train \
+  --dataset.repo_id=lerobot/libero \
+  --dataset.revision=a1aaacb7f6cd6ee5fb43120f673cebb0cfea7dd4 \
+  --dataset.video_backend=torchcodec \
+  --dataset.return_uint8=true \
+  --dataset.image_transforms.enable=true \
+  --dataset.use_imagenet_stats=true \
+  --dataset.eval_split=0.0 \
+  --policy.type=evo1 \
+  --policy.training_stage=stage1 \
+  --policy.apply_training_stage_defaults=true \
+  --policy.vlm_model_name="${VLM_DIR}" \
+  --policy.vlm_num_layers=14 \
+  --policy.vlm_dtype=bfloat16 \
+  --policy.device=cuda \
+  --policy.use_amp=true \
+  --policy.use_flash_attn=true \
+  --policy.enable_gradient_checkpointing=true \
+  --policy.gradient_checkpointing_use_reentrant=false \
+  --policy.image_resolution='[448,448]' \
+  --policy.chunk_size=50 \
+  --policy.n_action_steps=50 \
+  --policy.max_state_dim=24 \
+  --policy.max_action_dim=24 \
+  --policy.dropout=0.2 \
+  --policy.optimizer_lr=1e-5 \
+  --policy.optimizer_weight_decay=1e-3 \
+  --policy.optimizer_grad_clip_norm=1.0 \
+  --policy.scheduler_warmup_steps=1000 \
+  --policy.push_to_hub=false \
+  --use_policy_training_preset=true \
+  --batch_size=64 \
+  --steps=5000 \
+  --save_checkpoint=true \
+  --save_checkpoint_to_hub=false \
+  --save_freq=2500 \
+  --log_freq=10 \
+  --env_eval_freq=0 \
+  --num_workers=4 \
+  --prefetch_factor=2 \
+  --persistent_workers=true \
+  --seed=1000 \
+  --wandb.enable=false \
+  --output_dir=./outputs/evo1-libero-stage1-g128-5k
+```
+
+Stage 2 loads the Stage-1 policy but starts a fresh optimizer and scheduler. It trains for 80,000 steps;
+the reported checkpoint is the save at step 70,000:
+
+```bash
+uv run accelerate launch --num_processes=2 -m lerobot.scripts.lerobot_train \
+  --dataset.repo_id=lerobot/libero \
+  --dataset.revision=a1aaacb7f6cd6ee5fb43120f673cebb0cfea7dd4 \
+  --dataset.video_backend=torchcodec \
+  --dataset.return_uint8=true \
+  --dataset.image_transforms.enable=true \
+  --dataset.use_imagenet_stats=true \
+  --dataset.eval_split=0.0 \
+  --policy.path=./outputs/evo1-libero-stage1-g128-5k/checkpoints/005000/pretrained_model \
+  --policy.training_stage=stage2 \
+  --policy.apply_training_stage_defaults=true \
+  --policy.vlm_model_name="${VLM_DIR}" \
+  --policy.vlm_num_layers=14 \
+  --policy.vlm_dtype=float32 \
+  --policy.device=cuda \
+  --policy.use_amp=true \
+  --policy.use_flash_attn=true \
+  --policy.enable_gradient_checkpointing=true \
+  --policy.gradient_checkpointing_use_reentrant=false \
+  --policy.image_resolution='[448,448]' \
+  --policy.chunk_size=50 \
+  --policy.n_action_steps=50 \
+  --policy.max_state_dim=24 \
+  --policy.max_action_dim=24 \
+  --policy.dropout=0.2 \
+  --policy.optimizer_lr=1e-5 \
+  --policy.optimizer_weight_decay=1e-3 \
+  --policy.optimizer_grad_clip_norm=1.0 \
+  --policy.scheduler_warmup_steps=1000 \
+  --policy.push_to_hub=false \
+  --use_policy_training_preset=true \
+  --batch_size=64 \
+  --steps=80000 \
+  --resume=false \
+  --save_checkpoint=true \
+  --save_checkpoint_to_hub=false \
+  --save_freq=10000 \
+  --log_freq=10 \
+  --env_eval_freq=0 \
+  --num_workers=4 \
+  --prefetch_factor=2 \
+  --persistent_workers=true \
+  --seed=1000 \
+  --wandb.enable=false \
+  --output_dir=./outputs/evo1-libero-stage2-g128-80k
+```
+
+#### Author-format evaluation profile
+
+The author-format EVO1 LIBERO profile uses the raw LIBERO camera feature names
 (`observation.images.agentview_image` and `observation.images.robot0_eye_in_hand_image`), replans every
 14 actions, and binarizes the gripper command before stepping the simulator. The EVO1 policy postprocessor
 can crop the padded 24D action back to the 7D LIBERO action space and apply that gripper binarization. To
-evaluate a LIBERO checkpoint under the same one-episode-per-task setting, keep the raw camera names instead
-of the default `image`/`image2` mapping and set the LIBERO action postprocessing flags:
+evaluate an author-format checkpoint under the same one-episode-per-task setting, keep the raw camera names
+instead of the default `image`/`image2` mapping and set the LIBERO action postprocessing flags:

 ```bash
 lerobot-eval \
@@ -181,6 +319,75 @@ lerobot-eval \
  --eval.n_episodes=1
 ```

+#### Native `lerobot/libero` v3 profile
+
+Revision `a1aaacb7f6cd6ee5fb43120f673cebb0cfea7dd4` stores camera features as `image` and
+`image2`. This example evaluates all ten LIBERO Object tasks, launching each task in a fresh process:
+
+```bash
+export MUJOCO_GL=egl
+export PYOPENGL_PLATFORM=egl
+
+suite=libero_object
+horizon=280
+for task_id in {0..9}; do
+  lerobot-eval \
+    --policy.path=zuoxingdong/evo1_libero \
+    --policy.pretrained_revision=515921f4a2c1d3f3ad523721eafa26fdf2af315b \
+    --policy.vlm_model_name=OpenGVLab/InternVL3-1B-hf \
+    --policy.device=cuda \
+    --policy.use_amp=true \
+    --policy.vlm_dtype=bfloat16 \
+    --policy.use_flash_attn=false \
+    --policy.enable_gradient_checkpointing=false \
+    --policy.vlm_num_layers=14 \
+    --policy.image_resolution='[448,448]' \
+    --policy.max_text_length=1024 \
+    --policy.chunk_size=50 \
+    --policy.n_action_steps=14 \
+    --policy.max_state_dim=24 \
+    --policy.max_action_dim=24 \
+    --policy.num_inference_timesteps=32 \
+    --policy.postprocess_action_dim=7 \
+    --policy.binarize_gripper=true \
+    --policy.gripper_threshold=0.0 \
+    --policy.gripper_below_threshold_value=-1.0 \
+    --policy.gripper_above_threshold_value=1.0 \
+    --env.type=libero \
+    --env.task="${suite}" \
+    --env.task_ids="[${task_id}]" \
+    --env.camera_name=agentview_image,robot0_eye_in_hand_image \
+    --env.camera_name_mapping="{agentview_image: image, robot0_eye_in_hand_image: image2}" \
+    --env.control_mode=relative \
+    --env.obs_type=pixels_agent_pos \
+    --env.observation_width=448 \
+    --env.observation_height=448 \
+    --env.init_states=true \
+    --env.episode_length="${horizon}" \
+    --env.render_mode=rgb_array \
+    --env.max_parallel_tasks=1 \
+    --eval.n_episodes=50 \
+    --eval.batch_size=1 \
+    --eval.use_async_envs=false \
+    --eval.recording=false \
+    --seed=1000 \
+    --output_dir="./outputs/evo1-libero-stage2-70k-eval/${suite}/task-${task_id}" \
+    --job_name="evo1-libero-stage2-70k-${suite}-task-${task_id}"
+done
+```
+
+Run all ten task IDs for each suite with these horizons:
+
+| `env.task`       | `env.episode_length` |
+| ---------------- | -------------------: |
+| `libero_spatial` |                `280` |
+| `libero_object`  |                `280` |
+| `libero_goal`    |                `300` |
+| `libero_10`      |                `520` |
+
+Set `suite` and `horizon` for each row. This gives 500 episodes per suite and 2,000 episodes overall, while
+the loop's fresh process per task matches the measured RNG-reset topology.
+
 ## References

 - [EVO1 repository](https://github.com/MINT-SJTU/Evo-1)
@@ -62,7 +62,10 @@ Reference data points on a 4×H100 80 GB cluster (`accelerate launch --num_proce
 | `smolvla`   | 27m 49s    |      0.312 |           0.011 | ~80%     | `--policy.path=lerobot/smolvla_base`, `freeze_vision_encoder=false`, `train_expert_only=false`                                 |
 | `pi05`      | 3h 41m     |      2.548 |           0.014 | ~95%     | `--policy.pretrained_path=lerobot/pi05_base`, `gradient_checkpointing=true`, `dtype=bfloat16`, vision encoder + expert trained |

-The `dataloading_s` vs. `update_s` ratio is the diagnostic that matters: when `dataloading_s` approaches `update_s`, more GPUs stop helping — your dataloader is the bottleneck and you should look at `--num_workers`, image resolution, and disk speed before adding compute.
+Training logs separate the full iteration into `dataloading_s` (`next(dl_iter)`), `preprocessing_s`
+(image conversion and the policy pipeline), and `update_s` (the optimizer update). `step_s` covers all
+three and drives `samples_per_s`. The benchmark above predates this split, so its `dataloading_s` includes
+preprocessing.

 ### Schedule and checkpoints

@@ -30,7 +30,7 @@ The goal: lower the barrier to entry for robotics, so that everyone can contribu
 </div>

 <div align="center">
-  <img src="../../media/readme/robots_control_video.webp" width="640px" alt="Reachy 2 Demo">
+  <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/robots_control_video.webp" width="640px" alt="Reachy 2 Demo">
 </div>

 ## How It Works
@@ -241,24 +241,129 @@ See the [Real-Time Chunking](./rtc) guide for details on tuning RTC parameters.

 ---

+## Interactive Sessions
+
+Add `--interactive=true` to drive the rollout from the terminal instead of starting immediately. Hardware connects and the policy loads as usual, but **the robot stays still until you type `/start`** — useful when you want to position the scene first, re-instruct the policy between attempts, or run several takes without paying the load time again.
+
+```bash
+lerobot-rollout \
+    --strategy.type=base \
+    --policy.path=${HF_USER}/my_smolvla_policy \
+    --robot.type=so100_follower \
+    --robot.port=/dev/ttyACM0 \
+    --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+    --task="pick up the cube" \
+    --interactive=true
+```
+
+| Command           | Action                                                                                                                                                                             |
+| ----------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `/start`          | Start (or restart) the policy control loop                                                                                                                                         |
+| `/subtask <text>` | Change the instruction the policy follows, without stopping. No argument prints the current task. Only affects policies that condition on language (SmolVLA, π0/π0.5, and similar) |
+| `/reset`          | Stop movement, return the robot to its startup position, and restore the `--task` instruction                                                                                      |
+| `/stop`           | End the session and run the normal shutdown routines                                                                                                                               |
+| `/help`           | List the commands                                                                                                                                                                  |
+
+```text
+> /start
+Rollout running — task 'pick up the cube'. /subtask <text> to change it, ...
+> /subtask put the cube in the box
+Task: 'pick up the cube' → 'put the cube in the box' (applies from the next policy inference)
+> /reset
+Task restored to 'pick up the cube'
+Resetting — returning the robot to its initial position...
+Robot reset — holding at initial position. /start to run.
+> /stop
+```
+
+`Ctrl-C` still shuts down as usual, and closing stdin (`Ctrl-D`, or the end of a piped script) ends the session — so a piped script must keep stdin open for the intended duration:
+
+```bash
+(printf '/start\n'; sleep 60; printf '/stop\n') | lerobot-rollout ... --interactive=true
+```
+
+**How `/subtask` reaches the policy.** The stdin reader publishes the new instruction to the inference engine, which picks it up on its own inference thread, so nothing is mutated across threads while the robot is moving. How quickly the behavior changes depends on the backend:
+
+- **Sync** (`--inference.type=sync`) — precomputed chunk actions are dropped, so the new instruction applies on the very next control tick. Without this a chunking policy would keep executing up to `chunk_size` stale actions (seconds of the old behavior). Only the queued actions are discarded, so observation history and the rest of the episode state are preserved.
+- **RTC** (`--inference.type=rtc`) — the next chunk is generated under the new instruction and merged over the previous chunk's leftover prefix, so the switch lands within one inference and the motion stays continuous. The queue is deliberately not cleared: that would leave the robot without commands for a full inference latency. (With blending turned off via `--inference.rtc.enabled=false` the queued chunk drains first, so the switch lands up to one chunk later.)
+
+With `--use_torch_compile=true`, a switch whose instruction tokenizes to a different length can trigger a recompilation on the next forward pass, pausing inference for as long as the original warm-up took. Prefer leaving compilation off for sessions where you expect to re-instruct the policy often.
+
+**Logs below ERROR are muted while the session runs** so routine output doesn't interleave with what you're typing; errors and fatal inference failures still show, and normal logging resumes when the session ends. The gate is process-wide (it also withholds INFO/WARNING from any file handler you attached for the duration). Run without `--interactive` to watch the live log.
+
+Sessions work over SSH and on headless machines — the command reader uses the terminal (or pipe) directly and needs no display server.
+
+**Recording while interactive.** `--strategy.type=sentry` also supports `--interactive=true`: the session records continuously while you steer it. Each `/start`…`/reset` segment saves complete episodes plus one final partial episode, the dataset stays open until shutdown, and **frames are labeled with the live task** — a `/subtask` changes both the policy conditioning and the recorded label from the same frame onwards.
+
+```bash
+lerobot-rollout \
+    --strategy.type=sentry \
+    --policy.path=${HF_USER}/my_smolvla_policy \
+    --robot.type=so100_follower \
+    --robot.port=/dev/ttyACM0 \
+    --dataset.repo_id=${HF_USER}/rollout_cube_sessions \
+    --task="pick up the cube" \
+    --interactive=true
+```
+
+The other recording strategies (episodic, DAgger, highlight) are not supported: they bind their own keyboard controls, which would compete with the command prompt for the same terminal.
+
+### Programmatic control
+
+Everything the CLI session does is available as a library API: `RolloutController` exposes thread-safe `start()` / `reset()` / `stop()` / `set_task()` methods plus a `RolloutEvent` callback, with no stdin, printing, or log muting attached — embed it in your own application, network server, or notebook:
+
+```python
+from threading import Event, Thread
+
+from lerobot.rollout import (
+    LinkedEvent,
+    RolloutController,
+    RolloutEvent,
+    build_rollout_context,
+    create_strategy,
+)
+
+parent = Event()  # your application's shutdown signal
+ctx = build_rollout_context(cfg, LinkedEvent(parent))  # loads policy, connects robot
+strategy = create_strategy(cfg.strategy)
+strategy.setup(ctx)
+
+controller = RolloutController(strategy, ctx, on_event=print)  # or your own observer
+serve_thread = Thread(target=controller.serve)  # serve() blocks; run it where you like
+serve_thread.start()
+
+controller.start()                        # robot starts executing the policy
+controller.set_task("grab the red cube")  # re-instruct mid-run
+controller.reset()                        # stop movement, return home, stay warm
+controller.stop()                         # end serve()
+
+serve_thread.join()
+strategy.teardown(ctx)  # teardown stays with the caller
+```
+
+Set `play_sounds=False` in the config unless you want the vocal announcements, and note that `build_rollout_context` requires the shutdown event to be a `LinkedEvent` (the controller ends run segments through its local flag; your `parent` event still forces a full shutdown). `InteractiveSession` itself is a thin front-end over this controller — commands map 1:1 onto its methods.
+
+---
+
 ## Common Flags

-| Flag                              | Description                                                       | Default |
-| --------------------------------- | ----------------------------------------------------------------- | ------- |
-| `--policy.path`                   | **Required.** HF Hub model ID or local checkpoint path            | --      |
-| `--robot.type`                    | **Required.** Robot type (e.g. `so100_follower`, `koch_follower`) | --      |
-| `--robot.port`                    | Serial port for the robot                                         | --      |
-| `--robot.cameras`                 | Camera configuration (JSON dict)                                  | --      |
-| `--fps`                           | Control loop frequency                                            | 30      |
-| `--duration`                      | Run time in seconds (0 = infinite)                                | 0       |
-| `--device`                        | Torch device (`cpu`, `cuda`, `mps`)                               | auto    |
-| `--task`                          | Task description (used when no dataset is provided)               | --      |
-| `--display_data`                  | Stream telemetry to Rerun visualization                           | false   |
-| `--display_ip` / `--display_port` | Remote Rerun server address                                       | --      |
-| `--interpolation_multiplier`      | Action interpolation factor                                       | 1       |
-| `--use_torch_compile`             | Enable `torch.compile` for inference                              | false   |
-| `--resume`                        | Resume a previous recording session                               | false   |
-| `--play_sounds`                   | Vocal synthesis for events                                        | true    |
+| Flag                              | Description                                                                                                                                   | Default |
+| --------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------- | ------- |
+| `--policy.path`                   | **Required.** HF Hub model ID or local checkpoint path                                                                                        | --      |
+| `--robot.type`                    | **Required.** Robot type (e.g. `so100_follower`, `koch_follower`)                                                                             | --      |
+| `--robot.port`                    | Serial port for the robot                                                                                                                     | --      |
+| `--robot.cameras`                 | Camera configuration (JSON dict)                                                                                                              | --      |
+| `--fps`                           | Control loop frequency                                                                                                                        | 30      |
+| `--duration`                      | Run time in seconds (0 = infinite)                                                                                                            | 0       |
+| `--device`                        | Torch device (`cpu`, `cuda`, `mps`)                                                                                                           | auto    |
+| `--task`                          | Task description (used when no dataset is provided)                                                                                           | --      |
+| `--display_data`                  | Stream telemetry to Rerun visualization                                                                                                       | false   |
+| `--display_ip` / `--display_port` | Remote Rerun server address                                                                                                                   | --      |
+| `--interpolation_multiplier`      | Action interpolation factor                                                                                                                   | 1       |
+| `--interactive`                   | Chat-style stdin session (see [Interactive Sessions](#interactive-sessions)); the robot stays idle until `/start`. Base and sentry strategies | false   |
+| `--use_torch_compile`             | Enable `torch.compile` for inference                                                                                                          | false   |
+| `--resume`                        | Resume a previous recording session                                                                                                           | false   |
+| `--play_sounds`                   | Vocal synthesis for events                                                                                                                    | true    |

 ---

@@ -108,6 +108,7 @@ own binding plus a matching image block, e.g.

 ```yaml
 ask_vqa_top:
+  route: vqa
  bindings:
    vqa_query: "emitted_at(t, style=vqa, role=user, camera=observation.images.top)"
    vqa: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.top)"
@@ -127,7 +128,9 @@ ask_vqa_top:
      }
 ```

-Add one such sub-recipe per camera the dataset records.
+Add one such sub-recipe per camera the dataset records. The explicit
+`route: vqa` marker makes a matching sparse VQA annotation take precedence
+over normal weighted blend selection; component names are purely descriptive.

 ## Layer 3 — training format

@@ -141,7 +144,20 @@ sample["target_message_indices"]

 The renderer does not apply a tokenizer chat template. Policy processors decide how to serialize the messages for their backbone, which keeps the same dataset usable across SmolVLA, Pi0.5, and any future VLM that expects OpenAI-style chat messages.

+## Blends
+
+Blend recipes select one weighted sub-recipe deterministically from the sample index.
+`recipes/subtask_mem.yaml` trains the compact core blend — high-level subtask prediction, low-level execution, and memory. `recipes/subtask_mem_vqa_speech.yaml` is the fuller variant that also adds VQA and spoken interjection responses.
+
+`recipes/subtask_joint.yaml` demonstrates joint sequence training rather than a
+weighted blend. For the same sample, its assistant subtask is supervised with
+text cross-entropy on the `low_level` stream while action prediction remains
+active, matching the joint setup from the π0.5 paper. Enable
+`--policy.joint_subtask_conditioning=true` to use that subtask conditioning at inference.
+
 ## Graceful absence

-If both language columns are missing, `None`, or empty, `RenderMessagesStep` is a no-op.
-If an event-scoped branch is selected on a frame without the required event row, rendering returns `None`, allowing a loader to retry another sample.
+If both language columns are missing, `None`, or empty, `RenderMessagesStep` uses
+the task string as low-level supervision when available and otherwise leaves the
+sample unchanged. For an annotated sample, if no recipe branch applies and no
+task fallback exists, rendering returns `None`, allowing a loader to retry another sample.
@@ -142,6 +142,22 @@ repo_id = "yaak-ai/L2D-v3"
 dataset = StreamingLeRobotDataset(repo_id)  # streams directly from the Hub
 ```

+Datasets stored in an [HF Storage Bucket](https://huggingface.co/docs/hub/storage-buckets) (`hf://buckets/`) can be streamed the same way by passing `repo_type="bucket"`:
+
+```python
+dataset = StreamingLeRobotDataset("my-org/my-bucket", repo_type="bucket")
+```
+
+Both options are available in `lerobot-train` through `--dataset.streaming=true`, and `--dataset.repo_type=bucket` to stream from a bucket instead of a Hub dataset repo:
+
+```bash
+lerobot-train \
+  --dataset.repo_id=my-org/my-bucket \
+  --dataset.repo_type=bucket \
+  --dataset.streaming=true \
+  ...
+```
+
 <div style="display:flex; justify-content:center; gap:12px; flex-wrap:wrap;">
  <figure style="margin:0; text-align:center;">
    <img
@@ -92,6 +92,20 @@ LIBERO supports two control modes — `relative` (default) and `absolute`. Diffe
 --env.control_mode=relative   # or "absolute"
 ```

+### Reset performance
+
+By default, LeRobot preserves LIBERO's hard-reset behavior. With fixed initial
+states enabled, you can opt into soft resets to skip rebuilding the simulator
+model and renderer on every episode:
+
+```bash
+--env.init_states=true --env.hard_reset=false
+```
+
+Soft resets are faster but are not bit-identical to hard resets after the
+environment's settling steps, so camera observations and policy results may
+differ slightly. Use hard resets when reproducing benchmark results.
+
 ### Policy inputs and outputs

 **Observations:**
@@ -134,6 +134,20 @@ LIBERO-plus supports two control modes — `relative` (default) and `absolute`.
 --env.control_mode=relative   # or "absolute"
 ```

+### Reset performance
+
+By default, LeRobot preserves LIBERO's hard-reset behavior. With fixed initial
+states enabled, you can opt into soft resets to skip rebuilding the simulator
+model and renderer on every episode:
+
+```bash
+--env.init_states=true --env.hard_reset=false
+```
+
+Soft resets are faster but are not bit-identical to hard resets after the
+environment's settling steps, so camera observations and policy results may
+differ slightly. Use hard resets when reproducing benchmark results.
+
 ### Policy inputs and outputs

 **Observations:**
@@ -242,6 +242,17 @@ python src/lerobot/scripts/augment_dataset_quantile_stats.py \
  --repo-id=your_dataset
 ```

+Recording, resuming, and merging aggregate quantiles from per-episode summaries, so `meta/stats.json` ends up holding a conservative envelope (`min` for `q <= 50`, `max` for `q > 50`) rather than whole-dataset quantiles. To estimate the latter, scan every episode with a running histogram:
+
+```bash
+python src/lerobot/scripts/augment_dataset_quantile_stats.py \
+  --repo-id=your_dataset \
+  --overwrite \
+  --skip-images
+```
+
+`--skip-images` keeps the existing image statistics and avoids video decoding when only `STATE`/`ACTION` need recomputing, and `--root` reads a local dataset instead of the Hub. These values are histogram estimates, subject to discretization and rebinning error, so they can differ from the conservative ones — which changes MolmoAct2's normalized targets and therefore its loss scale. Statistics already saved inside an existing checkpoint are not affected.
+
 Alternatively, train MolmoAct2 with mean/std normalization:

 ```bash
@@ -1,28 +1,29 @@
 # Multi-GPU Training

-This guide shows you how to train policies on multiple GPUs using [Hugging Face Accelerate](https://huggingface.co/docs/accelerate).
+LeRobot trains on multiple GPUs through [Hugging Face Accelerate](https://huggingface.co/docs/accelerate). Three data-parallel layouts are supported:
+
+| Layout   | What it does                                                  | Config                                                  |
+| -------- | ------------------------------------------------------------- | ------------------------------------------------------- |
+| **DDP**  | Replicates the full model on every GPU                        | default on any multi-GPU launch                         |
+| **FSDP** | Shards parameters, gradients, and optimizer state across GPUs | `--parallelism.dp_shard=N`                              |
+| **HSDP** | Shards within groups of GPUs, replicates across groups        | `--parallelism.dp_replicate=R --parallelism.dp_shard=S` |

 ## Installation

-`accelerate` is included in the `training` extra. Install it with:
+`accelerate` is included in the `training` extra:

 ```bash
 pip install 'lerobot[training]'
 ```

-## Training with Multiple GPUs
+## Launching

-You can launch training in two ways:
+Distributed training can be launched through both `torchrun` and `accelerate launch`. Accelerate is used as a plain launcher: it does not manage the training configuration, and every distributed training setting lives in LeRobot's own config system.

-### Option 1: Without config (specify parameters directly)
-
-You can specify all parameters directly in the command without running `accelerate config`:
+With `torchrun`:

 ```bash
-accelerate launch \
-  --multi_gpu \
-  --num_processes=2 \
-  $(which lerobot-train) \
+torchrun --nproc-per-node=2 $(which lerobot-train) \
  --dataset.repo_id=${HF_USER}/my_dataset \
  --policy.type=act \
  --policy.repo_id=${HF_USER}/my_trained_policy \
@@ -31,32 +32,10 @@ accelerate launch \
  --wandb.enable=true
 ```

-**Key accelerate parameters:**
-
- `--multi_gpu`: Enable multi-GPU training
- `--num_processes=2`: Number of GPUs to use
- `--mixed_precision=fp16`: Use fp16 mixed precision (or `bf16` if supported)
-
-### Option 2: Using accelerate config
-
-If you prefer to save your configuration, you can optionally configure accelerate for your hardware setup by running:
+With `accelerate launch` (as a plain launcher):

 ```bash
-accelerate config
-```
-
-This interactive setup will ask you questions about your training environment (number of GPUs, mixed precision settings, etc.) and saves the configuration for future use. For a simple multi-GPU setup on a single machine, you can use these recommended settings:
-
- Compute environment: This machine
- Number of machines: 1
- Number of processes: (number of GPUs you want to use)
- GPU ids to use: (leave empty to use all)
- Mixed precision: fp16 or bf16 (recommended for faster training)
-
-Then launch training with:
-
-```bash
-accelerate launch $(which lerobot-train) \
+accelerate launch --num_processes=2 $(which lerobot-train) \
  --dataset.repo_id=${HF_USER}/my_dataset \
  --policy.type=act \
  --policy.repo_id=${HF_USER}/my_trained_policy \
@@ -65,116 +44,133 @@ accelerate launch $(which lerobot-train) \
  --wandb.enable=true
 ```

-## How It Works
+With no `--parallelism.*` flags, a multi-process launch runs plain DDP. Multi-node runs use the standard `torchrun --nnodes/--node-rank/--rdzv-endpoint` flags (or `accelerate launch --num_machines/--machine_rank/--main_process_ip`).

-When you launch training with accelerate:
+> [!WARNING]
+> Accelerate's YAML config files (`accelerate launch --config_file some.yaml`, `accelerate config`) are not supported. They configure the engine through environment variables, bypassing LeRobot's configuration system, so `train_config.json` would no longer describe the settings a run actually used. `lerobot-train` therefore refuses to start when [accelerate environment variables](https://huggingface.co/docs/accelerate/usage_guides/fsdp) are set. Put the settings in `--parallelism.*` / `--accelerator.*` flags instead, or set `LEROBOT_ALLOW_ACCELERATE_ENV=1` to acknowledge the override and proceed anyway.

-1. **Automatic detection**: LeRobot automatically detects if it's running under accelerate
-2. **Data distribution**: Your batch is automatically split across GPUs
-3. **Gradient synchronization**: Gradients are synchronized across GPUs during backpropagation
-4. **Single process logging**: Only the main process logs to wandb and saves checkpoints
+## Batch semantics, learning rate, and steps

-## Learning Rate and Training Steps Scaling
+Each of the `dp_replicate × dp_shard` data-parallel workers loads its own `--batch_size` micro-batch every step, so one training step consumes `batch_size × dp_world_size` samples, and `× gradient_accumulation_steps` of those go into each optimizer update:

-**Important:** LeRobot does **NOT** automatically scale learning rates or training steps based on the number of GPUs. This gives you full control over your training hyperparameters.
-
-### Why No Automatic Scaling?
-
-Many distributed training frameworks automatically scale the learning rate by the number of GPUs (e.g., `lr = base_lr × num_gpus`).
-However, LeRobot keeps the learning rate exactly as you specify it.
-
-### When and How to Scale
-
-If you want to scale your hyperparameters when using multiple GPUs, you should do it manually:
-
-**Learning Rate Scaling:**
-
-```bash
-# Example: 2 GPUs with linear LR scaling
-# Base LR: 1e-4, with 2 GPUs -> 2e-4
-accelerate launch --num_processes=2 $(which lerobot-train) \
-  --optimizer.lr=2e-4 \
-  --dataset.repo_id=lerobot/pusht \
-  --policy.type=act
+```
+effective_batch_size = batch_size × dp_world_size × gradient_accumulation_steps
 ```

-**Training Steps Scaling:**
+The training banner prints this factorization at startup. `--steps` counts loop steps (micro-batches per worker), not optimizer updates.

-Since the effective batch size `bs` increases with multiple GPUs (batch_size × num_gpus), you may want to reduce the number of training steps proportionally:
+Gradient accumulation is a first-class flag:

 ```bash
-# Example: 2 GPUs with effective batch size 2x larger
-# Original: batch_size=8, steps=100000
-# With 2 GPUs: batch_size=8 (16 in total), steps=50000
-accelerate launch --num_processes=2 $(which lerobot-train) \
-  --batch_size=8 \
-  --steps=50000 \
-  --dataset.repo_id=lerobot/pusht \
-  --policy.type=act
+torchrun --nproc-per-node=2 $(which lerobot-train) \
+  --batch_size=8 --accelerator.gradient_accumulation.steps=4 ...
 ```

-## Training Large Models with FSDP
+**LeRobot does not auto-scale the learning rate or the number of steps** when the effective batch size grows. If you scale out and want equivalent training, please adjust manually, e.g. with 2 GPUs: double `--optimizer.lr` (linear scaling), or halve `--steps`.

-DDP replicates the full model on every GPU, so a model that doesn't fit on one GPU won't fit under
-DDP either. For large models, use **FSDP** (Fully Sharded Data Parallel), which shards parameters,
-gradients, and optimizer state across GPUs. See the [accelerate FSDP guide](https://huggingface.co/docs/accelerate/usage_guides/fsdp) for background.
+## Sharded training (FSDP)

-An example on how to launch LeRobot training with FSDP across 4 GPUs (1 machine):
+If a model is too large to train with DDP, shard it with FSDP2:

 ```bash
-accelerate launch --config_file fsdp.yaml --num_processes=4 $(which lerobot-train) \
+torchrun --nproc-per-node=4 $(which lerobot-train) \
  --dataset.repo_id=${HF_USER}/my_dataset \
  --policy.type=<your_policy> \
+  --parallelism.dp_shard=4 \
+  --accelerator.mixed_precision=bf16 \
  --output_dir=outputs/train/my_policy_fsdp
 ```

-A minimal `fsdp.yaml` (FSDP1; shards params/grads/optimizer — ZeRO-3-equivalent):
+`--parallelism.dp_shard=-1` shards over however many processes the launcher started.

-```yaml
-compute_environment: LOCAL_MACHINE
-distributed_type: FSDP
-mixed_precision: bf16
-num_machines: 1
-num_processes: 4
-fsdp_config:
-  fsdp_version: 1
-  fsdp_sharding_strategy: FULL_SHARD # params + grads + optimizer (ZeRO-3)
-  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
-  fsdp_transformer_layer_cls_to_wrap: <YourTransformerBlock> # repeated block class to shard
-  fsdp_use_orig_params: true # required: optimizer is built pre-prepare
-  fsdp_state_dict_type: FULL_STATE_DICT
+### Wrap units
+
+FSDP shards the model in units (typically the repeated transformer block) and gathers one unit at a time during forward/backward. Policies declare their wrap units via `_fsdp_wrap_modules` on the policy class. For example, ACT declares `["ACTEncoderLayer", "ACTDecoderLayer"]` and FastWAM declares `["MoTLayer"]`. For a policy without a `_fsdp_wrap_modules` declaration, pass one of the flags below. You can specify the module class name explicitly, or use a size-based policy instead:
+
+```bash
+--accelerator.fsdp.wrap_modules='["MyTransformerBlock"]'   # explicit class names
+--accelerator.fsdp.min_num_params=1000000                  # or: wrap every submodule above 1M params
 ```

-Set `fsdp_transformer_layer_cls_to_wrap` to your model's repeated transformer-block class so each
-block is sharded as its own unit. `fsdp_use_orig_params: true` is required because LeRobot builds the
-optimizer before `accelerator.prepare()`.
+If a policy doesn't declare `_fsdp_wrap_modules` and no `--accelerator.fsdp.wrap_modules` or `--accelerator.fsdp.min_num_params` is passed, the run fails at startup rather than silently wrapping only the root module (which would forfeit all sharding memory savings).

-### FSDP checkpoints
+Other sharding settings:

-LeRobot gathers the full state dict across all ranks and the main process writes it as a single
-`model.safetensors`, loadable as usual with `Policy.from_pretrained(...)`. Two things to look out for:
+- `--accelerator.fsdp.reshard_after_forward`: whether to keep each unit's parameters resident after forward.
+- `--accelerator.fsdp.cpu_offload`: keeps parameters, gradients and optimizer states on CPU.
+- `--accelerator.fsdp.ignored_modules`: a regex of module paths to keep unsharded.

- **Checkpoints store fp32 weights.** Under mixed precision (`bf16`/`fp16`) FSDP keeps an fp32 master
-  copy, and the checkpoint saves it (~2× the bf16 size on disk) so training can resume consistently
-  with the fp32 optimizer state; `from_pretrained` casts back to the policy dtype on load. FSDP-specific
-  caveat: an fp32 checkpoint is materialized in full precision on the target device _before_ casting,
-  so loading it for inference on a tight GPU can OOM even when the bf16 model would fit — load on CPU
-  first, or cast `model.safetensors` to the deployment dtype offline.
- The sharded optimizer state is gathered into a full (world-size-independent) state dict and saved
-  alongside the model in the same `optimizer_state.safetensors` / `optimizer_param_groups.json`
-  format as single-GPU training, so **resume-from-checkpoint is supported** with `--resume=true`.
-  Resume reshards both the model and the optimizer state to the _current_ FSDP topology, so you can
-  resume an FSDP checkpoint on a different number of GPUs. Note that the data sampler is only
-  sample-exact when the world size and batch size match the original run (a warning is logged
-  otherwise); the optimizer/model state itself is unaffected.
+### HSDP
+
+Hybrid Sharded Data Parallel: parameters, gradients and optimizer states are sharded across `dp_shard` ranks, and that sharding is replicated `dp_replicate` times. Parameter all-gathers and gradient reduce-scatters stay inside a shard group; only the all-reduce that synchronizes the replicas crosses between groups. The two degrees must multiply to the world size:
+
+```bash
+# 16 GPUs = 2 nodes × 8: shard within each node, replicate across nodes
+torchrun --nnodes=2 --nproc-per-node=8 ... $(which lerobot-train) \
+  --parallelism.dp_replicate=2 --parallelism.dp_shard=8 ...
+```
+
+## Checkpoints
+
+Every checkpoint contains a `pretrained_model/` directory and a `training_state/` directory:
+
+```text
+005000/  # the training step at that checkpoint
+├── pretrained_model/
+│   ├── config.json  # policy config
+│   ├── train_config.json  # the full training config
+│   ├── model.safetensors  # full weights (checkpoint_format ∈ {safetensors, safetensors_dcp}, or any non-sharded run)
+│   ├── pytorch_model_fsdp_0/  # DCP weight shards (checkpoint_format ∈ {dcp, safetensors_dcp})
+│   ├── policy_preprocessor.json  # preprocessor config (when the run has a preprocessor)
+│   ├── policy_preprocessor_step_*.safetensors  # state of the stateful preprocessor steps
+│   ├── policy_postprocessor.json  # postprocessor config (when the run has a postprocessor)
+│   └── policy_postprocessor_step_*.safetensors  # state of the stateful postprocessor steps
+└── training_state/
+    ├── training_step.json  # step counter, topology, and batch semantics
+    ├── rng_state.safetensors  # rng states
+    ├── scheduler_state.json  # scheduler state (when the run has a scheduler)
+    ├── optimizer_state.safetensors  # full optimizer state (non-sharded runs)
+    ├── optimizer_param_groups.json  # optimizer param groups (non-sharded runs)
+    └── optimizer_0/  # DCP optimizer shards (sharded runs)
+```
+
+During single-GPU or DDP training, the pipeline serializes each state dict into a single file: `model.safetensors` for the model and `optimizer_state.safetensors` for the optimizer.
+
+During sharded training, the optimizer state is saved as DCP shards under `training_state/optimizer_0/`, and the layout of the model under `pretrained_model/` can be configured through `--checkpoint_format`:
+
+| `--checkpoint_format`     | Weights artifact                             | Use when                                                              |
+| ------------------------- | -------------------------------------------- | --------------------------------------------------------------------- |
+| `safetensors` _(default)_ | single `model.safetensors` only              | you want every checkpoint immediately loadable with `from_pretrained` |
+| `dcp`                     | `pytorch_model_fsdp_0/` shard directory only | gathering the full weights makes saves and resumes too slow           |
+| `safetensors_dcp`         | both                                         | you want fast resume _and_ immediately loadable checkpoints           |
+
+Two things to know about gathered (`safetensors`) checkpoints from sharded runs:
+
+- **They store fp32 weights.** Under mixed precision training, FSDP keeps an fp32 master copy, and the checkpoint saves the master copy to make sure training resumes consistently.
+- The gather is collective (all ranks participate) but only the main process writes.
+
+### Converting DCP checkpoints
+
+`lerobot-convert-dcp` merges a DCP shard directory into a regular `model.safetensors`, offline and without GPUs:
+
+```bash
+lerobot-convert-dcp --checkpoint_dir=outputs/train/run/checkpoints/005000
+lerobot-convert-dcp --checkpoint_dir=... --delete_dcp=true --push_to_hub=${HF_USER}/my_policy
+```
+
+`--push_to_hub` publishes the converted directory as a model repo.
+
+### Resuming
+
+Resume with `--resume=true --config_path=.../checkpoints/last/pretrained_model/train_config.json`. Resuming from a DCP checkpoint supports resharding the model and optimizer state to the _current_ topology, which means you can resume with a different `dp_replicate/dp_shard` split. The data sampler can always resume at the right epoch and offset, but is only _sample-exact_ when the world size and batch size match the original run (a warning is logged otherwise).
+
+> [!NOTE]
+> FSDP checkpoints written by LeRobot 0.6.x and earlier used a different on-disk layout (a gathered full optimizer state) and **cannot be resumed**.

 ## Notes

- The `--policy.use_amp` flag in `lerobot-train` is only used when **not** running with accelerate. When using accelerate, mixed precision is controlled by accelerate's configuration.
- Training logs, checkpoints, and hub uploads are only done by the main process to avoid conflicts. Non-main processes have console logging disabled to prevent duplicate output.
- The effective batch size is `batch_size × num_gpus`. If you use 4 GPUs with `--batch_size=8`, your effective batch size is 32.
- Learning rate scheduling is handled correctly across multiple processes—LeRobot sets `step_scheduler_with_optimizer=False` to prevent accelerate from adjusting scheduler steps based on the number of processes.
- When saving or pushing models, LeRobot automatically unwraps the model from accelerate's distributed wrapper to ensure compatibility.
- WandB integration automatically initializes only on the main process, preventing multiple runs from being created.
+- Checkpoint saves and end-of-training publishes are collective (every rank enters them). Gathered weights, sidecar files and Hub uploads are written by the main process alone.
+- Metrics are reduced across ranks before logging: losses are averaged, and `samples/s` reports cluster-wide throughput.
+- Learning-rate scheduling is stepped once per training step regardless of the number of processes (`step_scheduler_with_optimizer=False` is baked in).

-For more advanced configurations and troubleshooting, see the [Accelerate documentation](https://huggingface.co/docs/accelerate). If you want to learn more about how to train on a large number of GPUs, checkout this awesome guide: [Ultrascale Playbook](https://huggingface.co/spaces/nanotron/ultrascale-playbook).
+For background on the underlying machinery, see the [Accelerate FSDP guide](https://huggingface.co/docs/accelerate/usage_guides/fsdp). To go deeper on large-scale training, check out the [Ultrascale Playbook](https://huggingface.co/spaces/nanotron/ultrascale-playbook).
@@ -127,6 +127,17 @@ lerobot-edit-dataset \

 Or keep the dataset as-is and pass `--policy.normalization_mapping='{"ACTION": "MEAN_STD", "STATE": "MEAN_STD", "VISUAL": "IDENTITY"}'`.

+Recording, resuming, and merging aggregate quantiles from per-episode summaries, so `meta/stats.json` ends up holding a conservative envelope (`min` for `q <= 50`, `max` for `q > 50`) rather than whole-dataset quantiles. To estimate the latter, scan every episode with a running histogram:
+
+```bash
+python src/lerobot/scripts/augment_dataset_quantile_stats.py \
+    --repo-id=your_dataset \
+    --overwrite \
+    --skip-images
+```
+
+`--skip-images` keeps the existing image statistics and avoids video decoding when only `STATE`/`ACTION` need recomputing, and `--root` reads a local dataset instead of the Hub. These values are histogram estimates, subject to discretization and rebinning error, so they can differ from the conservative ones — which changes π₀.₅'s normalized targets and therefore its loss scale. Statistics already saved inside an existing checkpoint are not affected.
+
 ### Training Command Example

 The same finetune with the VLM frozen: less memory, at some cost in success rate. Swap `--dataset.repo_id` for your own dataset.
@@ -2,6 +2,25 @@

 https://diffusion-policy.cs.columbia.edu

+## Training
+
+The reference implementation maintains an exponential moving average (EMA) of the policy weights during training and evaluates the EMA weights. To reproduce this behavior, enable the trainer's EMA shadow:
+
+```bash
+lerobot-train \
+  --policy.type=diffusion \
+  --ema.enable=true \
+  ...
+```
+
+Checkpoints then contain a directly loadable copy of the EMA weights next to the live ones, e.g. for evaluation:
+
+```bash
+lerobot-eval --policy.path=outputs/train/.../checkpoints/last/pretrained_model_ema ...
+```
+
+The EMA decay schedule (`--ema.inv_gamma`, `--ema.power`, ...) defaults to the reference implementation's values. For a constant decay instead of the warmup schedule (e.g. to match openpi's pi0/pi05 training), set `--ema.decay=0.99`.
+
 ## Citation

 ```bibtex
@@ -59,6 +59,22 @@ When `use_relative_actions=true`, the training script automatically:

 ---

+## EMA of the policy weights
+
+OpenPI maintains an exponential moving average of the weights during training (`ema_decay=0.99` by default) and keeps the EMA copy for inference. To reproduce this with the LeRobot trainer, enable the EMA shadow with a constant decay:
+
+```bash
+python -m lerobot.scripts.lerobot_train \
+  --policy.type=pi05 \
+  --dataset.repo_id=your_org/your_dataset \
+  --ema.enable=true \
+  --ema.decay=0.99
+```
+
+Checkpoints then contain a directly loadable copy of the EMA weights in `pretrained_model_ema/` next to the live ones. Note that the shadow is a full extra copy of the parameters on the GPU. Like OpenPI (which disables EMA in its LoRA configs), EMA is not supported together with PEFT adapters.
+
+---
+
 ## Citation

 If you use this work, please cite both **OpenPI** and the π₀.₅ paper:
@@ -0,0 +1,339 @@
+# Third-Party Robots & Teleoperators
+
+The LeRobot ecosystem extends far beyond its officially supported hardware. Thanks to LeRobot's plugin architecture, the community has built integrations for a wide range of robot arms and teleoperation devices — from industrial manipulators to affordable hobbyist platforms, VR headsets, haptic devices, and full arm-plus-teleoperator kits. This page showcases community-maintained integrations you can use for teleoperation, data collection, and policy deployment.
+
+> [!IMPORTANT]
+> These projects are developed and maintained by third parties. Please refer to each repository for installation instructions, hardware requirements, and support.
+
+Drop-in plugins are auto-discovered by package name: LeRobot imports any installed package prefixed with `lerobot_robot_` or `lerobot_teleoperator_`. Once installed, reference the `type` the plugin registers (see its README — it may differ from the package name) directly from any LeRobot command:
+
+```bash
+pip install lerobot_robot_<name> lerobot_teleoperator_<name>
+
+lerobot-record \
+    --robot.type=<robot_name> \
+    --teleop.type=<teleoperator_name> \
+    --dataset.repo_id=${HF_USER}/my-dataset \
+    --dataset.num_episodes=5
+```
+
+> [!TIP]
+> ⚠️ marks projects that are forks/extensions of LeRobot. They may require custom setup rather than working with an unmodified install. All other entries are drop-in plugins.
+
+## Industrial & Collaborative Arms
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/SpesRobotics/lerobot-robot-xarm">lerobot-robot-xarm</a></td>
+      <td>Plugin for the xArm collaborative arm series from <a href="https://www.ufactory.cc/">UFACTORY</a>.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/lebai-robotics/lerobot_lebai">lerobot_lebai</a></td>
+      <td>Plugin for the six-axis collaborative arms from <a href="https://lebai.ltd/en/">Lebai</a>.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/wengmister/LeFranX">LeFranX</a> ⚠️</td>
+      <td>LeRobot extension for the <a href="https://franka.de/">Franka</a> research arm, paired with the <a href="https://www.robotera.com/">RobotEra XHand</a> hand for VR teleoperation.</td>
+    </tr>
+  </tbody>
+</table>
+
+#### Universal Robots UR5e
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/yechen056/UR5e-LeRobot">UR5e-LeRobot</a> ⚠️</td>
+      <td>LeRobot extension for the <a href="https://www.universal-robots.com/">Universal Robots UR5e</a>, with single-arm and bimanual support.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/scy-v/lerobot_ur5e_auto">lerobot_ur5e_auto</a> ⚠️</td>
+      <td>LeRobot extension for a mobile <a href="https://www.universal-robots.com/">Universal Robots UR5e</a>, adding automated recording at scale with minimal supervision.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/F-Fer/lerobot_ur5e_gello">lerobot_robot_ur5e</a></td>
+      <td>Plugin for the <a href="https://www.universal-robots.com/">Universal Robots UR5e</a> with a <a href="https://robotiq.com/">Robotiq</a> gripper, over RTDE control.</td>
+    </tr>
+  </tbody>
+</table>
+
+## Research & Learning Arms
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/TrossenRobotics/lerobot_trossen">lerobot_trossen</a></td>
+      <td>Plugin for the WidowX and ALOHA-style arms from <a href="https://www.trossenrobotics.com/">Trossen Robotics</a>.</td>
+    </tr>
+  </tbody>
+</table>
+
+#### AgileX Piper
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/AgRoboticsResearch/lerobot_robot_piper">lerobot_robot_piper (AgRobotics Research)</a></td>
+      <td>Plugin for the <a href="https://global.agilex.ai/">AgileX Piper</a> arm.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/WeGo-Robotics/lerobot_robot_piper">lerobot_robot_piper (WeGo Robotics)</a></td>
+      <td>Plugin for the <a href="https://global.agilex.ai/">AgileX Piper</a> arm, with multi-arm teleoperation, safety limits, and GUI tools.</td>
+    </tr>
+  </tbody>
+</table>
+
+## Affordable & Hobbyist Arms
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/servodevelop/fashionstar-lerobot-robot-cello">fashionstar-lerobot-robot-cello</a></td>
+      <td>Plugin for the StarAI Cello 6+1 degrees of freedom robot arm from <a href="https://fashionstar.com.hk/">FashionStar</a>.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/servodevelop/fashionstar-lerobot-robot-viola">fashionstar-lerobot-robot-viola</a></td>
+      <td>Plugin for the compact StarAI Viola 6+1 degrees of freedom robot arm from <a href="https://fashionstar.com.hk/">FashionStar</a>.</td>
+    </tr>
+  </tbody>
+</table>
+
+## Service, Mobile & Utility Robots
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/ugo-plus/lerobot-robot-ugo-pro">lerobot-robot-ugo-pro</a></td>
+      <td>Plugin for the ugo Pro dual-arm service robot from <a href="https://ugo.plus/products/ugo-pro/">ugo</a>.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/zuoxingdong/lerobot_robot_lekiwi_pincopen">lerobot_robot_lekiwi_pincopen</a></td>
+      <td>Plugin for a LeKiwi mobile manipulator with a <a href="https://github.com/pollen-robotics/PincOpen">PincOpen</a> gripper.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/KillingJacky/lerobot-robot-dummy">lerobot-robot-dummy</a></td>
+      <td>Plugin simulating a robot for recording without hardware. Useful for debugging !</td>
+    </tr>
+  </tbody>
+</table>
+
+## Teleoperators
+
+### VR & Motion Controllers
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/SpesRobotics/lerobot-teleoperator-teleop">lerobot-teleoperator-teleop</a></td>
+      <td>Plugin turning a phone or VR headset into a teleoperator via <a href="https://immersiveweb.dev">WebXR</a>, wrapping the open-source <a href="https://github.com/SpesRobotics/teleop"><code>teleop</code></a> library.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/Jas000n/lerobot-teleoperator-spacemouse">lerobot-teleoperator-spacemouse</a></td>
+      <td>Plugin for the <a href="https://3dconnexion.com/">3Dconnexion SpaceMouse</a>, with inverse kinematics for SO-ARMS robots.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/Dream-Machines-Robotics/vr-teleop-kit">vr-teleop-kit</a></td>
+      <td>Plugin teleoperating arms from a <a href="https://www.meta.com/quest/">Meta Quest</a> (WebXR), relying on URDF descriptions for inverse kinematics.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/xensedyl/lerobot-teleoperator-pico4">lerobot-teleoperator-pico4</a></td>
+      <td>Plugin for the <a href="https://www.picoxr.com/">PICO 4</a> VR headset, with a companion controller-free <a href="https://github.com/xensedyl/lerobot-teleoperator-pico4-hand">hand-tracking variant</a>.</td>
+    </tr>
+  </tbody>
+</table>
+
+### Leader Arms
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/F-Fer/lerobot_ur5e_gello">lerobot_teleoperator_gello</a></td>
+      <td>Plugin for the 7 degrees of freedom <a href="https://wuphilipp.github.io/gello_site/">GELLO</a> teleoperator.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/uynitsuj/lerobot_teleoperator_yamactiveleader">lerobot_teleoperator_yamactiveleader</a></td>
+      <td>Plugin for the active YAM teleoperator from <a href="https://i2rt.com/">I2RT</a>, a bilateral force-feedback arm.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/charlie8612/lerobot_teleoperator_omy">lerobot_teleoperator_omy</a></td>
+      <td>Plugin for the OMY-L100 6 degrees of freedom teleoperator from <a href="https://www.robotis.com/">ROBOTIS</a>.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://pypi.org/project/lerobot-teleoperator-pipermate/">lerobot-teleoperator-pipermate</a></td>
+      <td>Plugin for the PiperMate teleoperator (<a href="https://fashionstar.com.hk/">FashionStar</a> UART servos), driving the <a href="https://global.agilex.ai/">AgileX Piper</a> arm.</td>
+    </tr>
+  </tbody>
+</table>
+
+### Haptic Devices
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/chohh7391/lerobot_teleoperator_inverse3">lerobot_teleoperator_inverse3</a></td>
+      <td>Plugin for the <a href="https://www.haply.co/">Haply Inverse3</a> haptic device, adding force-feedback teleoperation.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/hzhz112/lerobot_teleoperator_omega7">lerobot_teleoperator_omega7</a></td>
+      <td>Plugin for the <a href="https://www.forcedimension.com/">Force Dimension omega.7</a> haptic device, adding force-feedback teleoperation.</td>
+    </tr>
+  </tbody>
+</table>
+
+### Networked & Remote
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://pypi.org/project/lerobot-teleoperator-livekit/">lerobot-teleoperator-livekit</a></td>
+      <td>Plugin receiving teleoperation commands over a <a href="https://livekit.io/">LiveKit</a> Portal (WebRTC) for remote control.</td>
+    </tr>
+  </tbody>
+</table>
+
+## Full Kits (Robot + Teleoperator)
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/villekuosmanen/lerobot-arx5">lerobot-arx5</a></td>
+      <td>Plugin for the <a href="https://www.arx-x.com/">ARX5</a> arm: <a href="https://pypi.org/project/lerobot-robot-arx5/"><code>lerobot-arx5</code></a> robot arm with its <a href="https://pypi.org/project/lerobot-teleoperator-arx5/"><code>lerobot-teleoperator-arx5</code></a> teleoperator arm.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/robertorobotics/Nextis-AIRA-3D">Nextis-AIRA-3D</a></td>
+      <td>Plugin for the 7 degrees of freedom arm from <a href="https://www.nextis.tech">Nextis</a>: robot arm <code>aira_follower</code> and teleoperator arm <code>aira_leader</code>.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/pravsels/lerobot_yam">lerobot_yam</a></td>
+      <td>Plugin suite for the YAM arm from <a href="https://i2rt.com/">I2RT</a>: robot arm <code>yam_follower</code> and teleoperator arm <code>yam_leader</code>.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/robot-learning-co/trlc-dk1">trlc-dk1</a></td>
+      <td>Plugin for the development kit from <a href="https://www.robot-learning.co/">The Robot Learning Company</a>: single and bimanual arms follower/teleoperator types.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/hexfellow/hex_lerobot_drivers">hex_lerobot_drivers</a></td>
+      <td>Plugin suite for <a href="https://hexfellow.com/">HEXFELLOW</a> devices: robots, teleoperators, and cameras (see <a href="./third_party_sensors">Cameras &amp; Sensors</a>).</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/Hiwonder-official/lerobot-robot-nexarm-follower">lerobot-robot-nexarm-follower</a></td>
+      <td>Plugin for the NexArm from <a href="https://www.hiwonder.com/">Hiwonder</a>: the <a href="https://github.com/Hiwonder-official/lerobot-robot-nexarm-follower">robot arm</a> and its matching <a href="https://github.com/Hiwonder-official/lerobot-teleoperator-nexarm-leader">teleoperator arm</a>.</td>
+    </tr>
+  </tbody>
+</table>
+
+## ROS 2 Bridges
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/ngres/leros2">leros2</a></td>
+      <td>Plugin bridging ROS 2 topics and actions to LeRobot robots and teleoperators.</td>
+    </tr>
+    <tr style="border:0">
+      <td><a href="https://github.com/ROBOTIS-GIT/lerobot_robot_ros2_zenoh">lerobot_robot_ros2_zenoh</a></td>
+      <td>Plugin bridging ROS 2 robots to LeRobot over <a href="https://zenoh.io">Zenoh</a> pub/sub transport.</td>
+    </tr>
+  </tbody>
+</table>
+
+## Contributing
+
+Built your own LeRobot hardware integration? The plugin system makes it straightforward to add new robots and teleoperators — check out the [Bring Your Own Hardware](./integrate_hardware) guide to get started, and share your project with the community!
@@ -0,0 +1,99 @@
+# Third-Party Cameras & Sensors
+
+The LeRobot ecosystem extends far beyond its natively supported cameras (OpenCV, Intel RealSense, ZMQ, Reachy 2). Thanks to LeRobot's plugin architecture, the community has built drop-in camera and sensor integrations — from depth cameras to vision-based tactile sensors. This page showcases community-maintained camera and sensor integrations you can use for teleoperation, data collection, and policy deployment.
+
+> [!IMPORTANT]
+> These projects are developed and maintained by third parties. Please refer to each repository for installation instructions, hardware requirements, and support.
+
+Drop-in plugins are auto-discovered by package name: LeRobot imports any installed package prefixed with `lerobot_camera_`. Once installed, reference the camera `type` the plugin registers (see its README — it may differ from the package name) directly from any LeRobot command:
+
+```bash
+pip install lerobot_camera_<name>
+
+lerobot-record \
+    --robot.type=so101_follower \
+    --robot.port=/dev/ttyACM0 \
+    --robot.cameras="{ front: {type: <name>, width: 640, height: 480, fps: 30} }" \
+    --dataset.repo_id=${HF_USER}/my-dataset \
+    --dataset.num_episodes=5
+```
+
+## Tactile Sensors
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/xensedyl/lerobot-camera-xense">lerobot-camera-xense</a></td>
+      <td>Plugin for <a href="https://www.xenserobotics.com/">Xense</a> vision-based tactile sensors, exposing rectified/difference images, depth, and 2D markers.</td>
+    </tr>
+  </tbody>
+</table>
+
+## Depth Cameras
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/hexfellow/hex_lerobot_drivers/tree/main/lerobot_camera_berxel">lerobot_camera_berxel</a></td>
+      <td>Plugin for the <a href="https://www.berxel.com/">Berxel</a> depth camera, part of the broader <a href="https://hexfellow.com/">HEXFELLOW</a> driver suite.</td>
+    </tr>
+  </tbody>
+</table>
+
+## Networked Cameras
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/F-Fer/lerobot_ur5e_gello">lerobot_camera_zmq</a></td>
+      <td>Plugin streaming <a href="https://www.stereolabs.com/">Stereolabs ZED</a> and USB camera frames from a Raspberry Pi over the network.</td>
+    </tr>
+  </tbody>
+</table>
+
+## Virtual Cameras
+
+<!-- prettier-ignore -->
+<table width="100%" style="display:table; width:100%; table-layout:fixed;">
+  <colgroup>
+    <col width="30%" />
+    <col width="70%" />
+  </colgroup>
+  <thead style="border:0">
+    <tr style="border:0"><th>Project</th><th>Description</th></tr>
+  </thead>
+  <tbody>
+    <tr style="border:0">
+      <td><a href="https://github.com/hexfellow/hex_lerobot_drivers/tree/main/lerobot_camera_dummy">lerobot_camera_dummy</a></td>
+      <td>Plugin simulating a camera for recording without hardware. Useful for debugging !</td>
+    </tr>
+  </tbody>
+</table>
+
+## Contributing
+
+Built your own LeRobot camera or sensor integration? Package it as an installable `lerobot_camera_<name>` plugin and it will be auto-discovered by the LeRobot CLI — see the [Bring Your Own Hardware](./integrate_hardware) guide and the [Cameras](./cameras) reference to get started, then share your project with the community!
@@ -40,3 +40,15 @@ lerobot-eval \
 ```

 However, in most cases, presence of an accelerator is detected automatically and `policy.device` parameter can be omitted from CLI commands.
+
+## Mixed precision
+
+Training precision is owned by `--accelerator.mixed_precision`, which accepts `no` (default) and `bf16`:
+
+```bash
+lerobot-train \
+    --policy.type=act \
+    --accelerator.mixed_precision=bf16 ...
+```
+
+`bf16` requires an accelerator that supports it.
@@ -30,19 +30,10 @@ Only Qwen + the action head are used. The world model is not needed at inference

 Available presets via `action_model_type`:

-| Preset  | Heads | Head dim |
-| ------- | ----- | -------- |
-| `DiT-B` | 12    | 64       |
-| `DiT-L` | 32    | 48       |
-
-The preset only sets the attention geometry, and each entry can be overridden by
-`action_num_heads` / `action_attention_head_dim`. Two widths follow from it:
-
- the DiT's **internal** width is `heads x head_dim` (768 for `DiT-B`), derived rather than configured;
- the DiT's **output** width, and the width of the action-decoder and state-encoder MLPs, is
-  `action_hidden_size` (default 1024).
-
-So `DiT-B` runs a 768-wide transformer that projects to 1024. The two are independent.
+| Preset  | Hidden dim | Heads | Head dim |
+| ------- | ---------- | ----- | -------- |
+| `DiT-B` | 768        | 12    | 64       |
+| `DiT-L` | 1536       | 32    | 48       |

 ### World model details

@@ -83,27 +74,10 @@ Key parameters in `VLAJEPAConfig`:
 | `num_inference_timesteps` | 4       | Euler integration steps for action denoising                                                                                                                                    |
 | `freeze_qwen`             | `False` | Freeze the Qwen3-VL backbone and only train the action head                                                                                                                     |
 | `reinit_modules`          | `None`  | Key prefixes allowed to be randomly re-initialised on load (for cross-embodiment transfer, see [Fine-tuning on a different embodiment](#fine-tuning-on-a-different-embodiment)) |
-| `resize_images_to`        | `None`  | `(height, width)` every camera frame is resized to before the Qwen3-VL vision tower. `None` keeps the native resolution, and Qwen3-VL's patch count grows with it, so a 720x1280 camera can exhaust GPU memory. The published checkpoints use `[224, 224]` |
-| `gripper_dim`             | 6       | Index of the gripper dimension in the action vector. Ignored when `gripper_joint_names` matches a dataset action name                                                            |
-| `gripper_joint_names`     | `["gripper"]` | Action-dimension names identifying the gripper; the matched index wins over `gripper_dim`                                                                                  |
-| `gripper_threshold`       | 0.5     | Threshold used by `pre_snap_gripper_action` and `binarize_gripper_action`. Note `binarize` runs *after* unnormalization, so this is compared against the gripper's physical value |
-| `pre_snap_gripper_action` | `False` | Snap the gripper dim to {0, 1} before unnormalization. LIBERO-specific, see below                                                                                                |
-| `binarize_gripper_action` | `False` | Binarize the gripper dim to {-1, 1} after unnormalization. LIBERO-specific, see below                                                                                            |
-| `clip_normalized_actions` | `True`  | Clip normalized actions to [-1, 1] before unnormalizing. Only applied when `ACTION` uses `MIN_MAX`; ignored (with a warning) under `MEAN_STD`, where it would truncate at 1 sigma |
-| `world_model_num_views`   | `None`  | Camera views the world-model predictor is built for. Baked into checkpoint shapes. `None` falls back to `jepa_tubelet_size`, which is what the published checkpoints encode       |
-
-<Tip warning={true}>
-
-`pre_snap_gripper_action` and `binarize_gripper_action` are a port of the starVLA LIBERO eval
-loop and are only correct for LIBERO's action convention. `pre_snap` writes {0, 1} into
-*normalized* space, the unnormalizer maps those to the midpoint and the max, and `binarize` then
-compares that **physical** value against `gripper_threshold` (0.5). For a gripper measured in
-degrees, mm or [0, 100], both values land above the threshold and the commanded gripper becomes a
-constant. They default to `False` for that reason; enable them only for LIBERO-style setups, and
-set `gripper_threshold` in the gripper's own units if you do. The processor factory warns when the
-dataset stats show the range cannot work.
-
-</Tip>
+| `gripper_dim`             | 6       | Index of the gripper dimension in the action vector (e.g. 6 for a 7-DoF arm with gripper as the last joint)                                                                     |
+| `gripper_threshold`       | 0.5     | Threshold used by `pre_snap_gripper_action` and `binarize_gripper_action` to binarize the gripper dimension                                                                     |
+| `pre_snap_gripper_action` | `True`  | Snap the gripper dim to {0, 1} before unnormalization. Set to `False` for robots without a binary gripper                                                                       |
+| `binarize_gripper_action` | `True`  | Binarize the gripper dim to {-1, 1} after unnormalization. Set to `False` for robots without a binary gripper                                                                   |

 ---

@@ -213,20 +187,14 @@ lerobot-eval \

 ## Fine-tuning on datasets with a different number of cameras

-The pretrained world model predictor was trained with `embed_dim = world_model_num_views × 1024`, i.e. two camera views.
-
-<Tip>
-
-This view count used to be read from `jepa_tubelet_size`, which also names the JEPA encoder's *temporal* tubelet size. `world_model_num_views` is the field for it now; leaving it at `None` falls back to `jepa_tubelet_size` so the published checkpoints keep loading unchanged.
-
-</Tip>
+The pretrained world model predictor was trained with `embed_dim = jepa_tubelet_size × 1024` (default `jepa_tubelet_size=2`).

 **Default behaviour — view padding / trimming (no action required)**

-When fine-tuning from `VLA-JEPA-Pretrain` the model automatically adjusts the number of views fed to the world model to match `world_model_num_views`:
+When fine-tuning from `VLA-JEPA-Pretrain` the model automatically adjusts the number of views fed to the world model to match `jepa_tubelet_size`:

 - **Single-view datasets (e.g. BridgeV2):** the single-view latent is duplicated to produce a two-view world-model input, preserving the JEPA self-supervised signal without any weight mismatch.
- **>2-view datasets (e.g. DROID with 3 views):** all views are passed to the Qwen backbone (for richer context), but only the first `world_model_num_views` views (one wrist + one third-person, following the configured view order) are used for the world model.
+- **>2-view datasets (e.g. DROID with 3 views):** all views are passed to the Qwen backbone (for richer context), but only the first `jepa_tubelet_size` views (one wrist + one third-person, following the configured view order) are used for the world model.

 **Option 1 — Disable the world model**

@@ -242,7 +210,7 @@ lerobot-train \

 **Option 2 — Reinitialize the predictor input projection**

-If you want to change `world_model_num_views` to a value other than 2, load the checkpoint with `strict=False` and reinitialize `model.video_predictor.predictor_embed` for the new `embed_dim`. All other predictor block weights (attention, MLP, norm, output projection) are camera-count-agnostic and can be reused from the pretrained checkpoint.
+If you want to change `jepa_tubelet_size` to a value other than 2, load the checkpoint with `strict=False` and reinitialize `model.video_predictor.predictor_embed` for the new `embed_dim`. All other predictor block weights (attention, MLP, norm, output projection) are camera-count-agnostic and can be reused from the pretrained checkpoint.

 ---

@@ -25,7 +25,7 @@ discord = "https://discord.gg/s3KuuzsPFb"

 [project]
 name = "lerobot"
-version = "0.6.1"
+version = "0.6.2"
 description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
 dynamic = ["readme"]
 license = { text = "Apache-2.0" }
@@ -87,7 +87,7 @@ dependencies = [

    # Build tools (required by opencv-python-headless on some platforms)
    "cmake>=3.29.0.1,<4.2.0",
-    "setuptools>=71.0.0,<81.0.0",
+    "setuptools>=71.0.0,<82.0.0",  # torch 2.11 requires setuptools<82; a higher cap makes the resolver downgrade torch
 ]

 # Optional dependencies
@@ -261,7 +261,7 @@ annotations = [
 # Development
 dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools>=1.73.1,<2.0.0", "mypy>=1.19.1", "ruff>=0.14.1", "lerobot[notebook]"]
 notebook = ["jupyter>=1.0.0,<2.0.0", "ipykernel>=6.0.0,<7.0.0"]
-test = ["pytest>=8.1.0,<9.0.0", "pytest-timeout>=2.4.0,<3.0.0", "pytest-cov>=5.0.0,<8.0.0", "mock-serial>=0.0.1,<0.1.0 ; sys_platform != 'win32'"]
+test = ["pytest>=8.1.0,<10.0.0", "pytest-timeout>=2.4.0,<3.0.0", "pytest-cov>=5.0.0,<8.0.0", "mock-serial>=0.0.1,<0.1.0 ; sys_platform != 'win32'"]
 video_benchmark = ["scikit-image>=0.23.2,<0.26.0", "pandas>=2.2.2,<2.4.0"]

 # Simulation
@@ -346,6 +346,7 @@ lerobot-record="lerobot.scripts.lerobot_record:main"
 lerobot-replay="lerobot.scripts.lerobot_replay:main"
 lerobot-setup-motors="lerobot.scripts.lerobot_setup_motors:main"
 lerobot-teleoperate="lerobot.scripts.lerobot_teleoperate:main"
+lerobot-convert-dcp="lerobot.scripts.lerobot_convert_dcp:main"
 lerobot-eval="lerobot.scripts.lerobot_eval:main"
 lerobot-train="lerobot.scripts.lerobot_train:main"
 lerobot-train-tokenizer="lerobot.scripts.lerobot_train_tokenizer:main"
@@ -475,6 +476,12 @@ default.extend-ignore-identifiers-re = [
 # TODO: Enable mypy gradually module by module across multiple PRs
 # Uncomment [tool.mypy] first, then uncomment individual module overrides as they get proper type annotations

+[tool.pytest.ini_options]
+markers = [
+    "multigpu: distributed tests needing 2-4 GPUs (CI: docker_publish.yml lane)",
+    "multigpu_heavy: 8-GPU sweeps and soak tests; never run in CI",
+]
+
 [tool.mypy]
 python_version = "3.12"
 ignore_missing_imports = true
@@ -521,6 +528,15 @@ disallow_untyped_defs = true
 disallow_incomplete_defs = true
 check_untyped_defs = true

+[[tool.mypy.overrides]]
+module = "lerobot.distributed.*"
+ignore_errors = false
+
+# extra strictness for the distributed engine
+disallow_untyped_defs = true
+disallow_incomplete_defs = true
+check_untyped_defs = true
+
 [[tool.mypy.overrides]]
 module = "lerobot.optim.*"
 ignore_errors = false
@@ -109,6 +109,11 @@ class RealSenseCamera(Camera):
        ```
    """

+    # Maximum number of warmup attempts made by connect(). A failed attempt is first
+    # retried with a plain pipeline stop/start, which is usually enough to recover the
+    # stream; a USB hardware reset is performed before the final attempt as a last resort.
+    _MAX_CONNECT_ATTEMPTS = 3
+
    def __init__(self, config: RealSenseCameraConfig):
        """
        Initializes the RealSenseCamera instance.
@@ -173,6 +178,76 @@ class RealSenseCamera(Camera):
        """Checks if the camera pipeline is started and streams are active."""
        return self.rs_pipeline is not None and self.rs_profile is not None

+    def _hardware_reset(self, wait_s: float = 5.0) -> None:
+        """Issue a USB hardware reset to recover an unresponsive device (common on D405)."""
+        context = rs.context()
+        for device in context.query_devices():
+            if device.get_info(rs.camera_info.serial_number) == self.serial_number:
+                logger.info(f"{self} performing hardware reset.")
+                device.hardware_reset()
+                time.sleep(wait_s)
+                return
+        logger.warning(f"{self} device not found for hardware reset, skipping.")
+
+    def _open_pipeline(self) -> None:
+        """Initializes the RealSense pipeline, starts it, and starts the background read thread.
+
+        Raises:
+            ValueError: If the configuration is invalid, a requested sensor option is unsupported,
+                or a requested sensor value is invalid.
+            ConnectionError: If the camera is found but fails to start the pipeline or no RealSense devices are detected at all.
+            RuntimeError: If the pipeline starts but fails to apply requested settings.
+        """
+        rs_pipeline = rs.pipeline()
+        rs_config = rs.config()
+        self._configure_rs_pipeline_config(rs_config)
+
+        try:
+            rs_profile = rs_pipeline.start(rs_config)
+        except RuntimeError as e:
+            raise ConnectionError(
+                f"Failed to open {self}.Run `lerobot-find-cameras realsense` to find available cameras."
+            ) from e
+
+        self.rs_pipeline = rs_pipeline
+        self.rs_profile = rs_profile
+
+        try:
+            self._configure_capture_settings()
+            self._configure_sensor_options()
+            self._start_read_thread()
+        except BaseException:
+            self._release_after_failed_setup()
+            raise
+
+    def _run_warmup(self) -> None:
+        """Blocks until at least one valid frame has been captured by the background thread.
+
+        Raises:
+            ConnectionError: If no frame arrives before ``warmup_s`` elapses.
+        """
+        # NOTE(Steven/Caroline): Enforcing at least one second of warmup as RS cameras need a bit of time before the first read. If we don't wait, the first read from the warmup will raise.
+        self.warmup_s = max(self.warmup_s, 1)
+
+        warmup_read = self.async_read if self.use_rgb else self.async_read_depth
+        start_time = time.time()
+        while time.time() - start_time < self.warmup_s:
+            warmup_read(timeout_ms=self.warmup_s * 1000)
+            time.sleep(0.1)
+        with self.frame_lock:
+            if (self.use_rgb and self.latest_color_frame is None) or (
+                self.use_depth and self.latest_depth_frame is None
+            ):
+                raise ConnectionError(f"{self} failed to capture frames during warmup.")
+
+    def _release_after_failed_setup(self) -> None:
+        """Releases the device handle and restores auto-detected settings after a failed attempt."""
+        try:
+            self._cleanup_resources()
+        except Exception:
+            logger.exception(f"Failed to fully clean up {self} after connect() failed.")
+        self._reset_connection_settings()
+
    @check_if_already_connected
    def connect(self, warmup: bool = True) -> None:
        """
@@ -181,58 +256,53 @@ class RealSenseCamera(Camera):
        Initializes the RealSense pipeline, configures the required streams (color
        and optionally depth), starts the pipeline, and validates the actual stream settings.

+        If the pipeline starts but no frames arrive during warmup, retries up to
+        ``_MAX_CONNECT_ATTEMPTS`` times, performing a USB hardware reset before the
+        final attempt.
+
        Args:
            warmup (bool): If True, waits at connect() time until at least one valid frame
                           has been captured by the background thread. Defaults to True.

        Raises:
            DeviceAlreadyConnectedError: If the camera is already connected.
-            ValueError: If the configuration is invalid, a requested sensor option is unsupported,
-                or a requested sensor value is invalid.
+            ValueError: If the configuration is invalid (e.g., missing serial/name, name not unique).
            ConnectionError: If the camera is found but fails to start the pipeline or no RealSense devices are detected at all.
            RuntimeError: If the pipeline starts but fails to apply requested settings.
        """

-        self.rs_pipeline = rs.pipeline()
-        rs_config = rs.config()
-        self._configure_rs_pipeline_config(rs_config)
+        if not warmup:
+            self._open_pipeline()
+            logger.info(f"{self} connected.")
+            return

-        try:
-            self.rs_profile = self.rs_pipeline.start(rs_config)
-        except RuntimeError as e:
-            self.rs_profile = None
-            self.rs_pipeline = None
-            raise ConnectionError(
-                f"Failed to open {self}.Run `lerobot-find-cameras realsense` to find available cameras."
-            ) from e
+        last_error: Exception | None = None

-        try:
-            self._configure_capture_settings()
-            self._configure_sensor_options()
-            self._start_read_thread()
+        for attempt in range(1, self._MAX_CONNECT_ATTEMPTS + 1):
+            if attempt == self._MAX_CONNECT_ATTEMPTS:
+                self._hardware_reset()

-            # NOTE(Steven/Caroline): Enforcing at least one second of warmup as RS cameras need a bit of time before the first read. If we don't wait, the first read from the warmup will raise.
-            self.warmup_s = max(self.warmup_s, 1)
+            self._open_pipeline()

-            warmup_read = self.async_read if self.use_rgb else self.async_read_depth
-            start_time = time.time()
-            while time.time() - start_time < self.warmup_s:
-                warmup_read(timeout_ms=self.warmup_s * 1000)
-                time.sleep(0.1)
-            with self.frame_lock:
-                if (self.use_rgb and self.latest_color_frame is None) or (
-                    self.use_depth and self.latest_depth_frame is None
-                ):
-                    raise ConnectionError(f"{self} failed to capture frames during warmup.")
-        except BaseException:
+            connected = False
            try:
-                self._cleanup_resources()
-            except Exception:
-                logger.exception(f"Failed to fully clean up {self} after connect() failed.")
-            self._reset_connection_settings()
-            raise
+                self._run_warmup()
+                connected = True
+            except (TimeoutError, ConnectionError) as e:
+                last_error = e
+            finally:
+                if not connected:
+                    self._release_after_failed_setup()

-        logger.info(f"{self} connected.")
+            if connected:
+                logger.info(f"{self} connected.")
+                return
+
+            logger.warning(f"{self} warmup failed (attempt {attempt}/{self._MAX_CONNECT_ATTEMPTS}).")
+
+        raise ConnectionError(
+            f"{self} failed to capture frames after {self._MAX_CONNECT_ATTEMPTS} attempts."
+        ) from last_error

    @staticmethod
    def find_cameras() -> list[dict[str, Any]]:
@@ -629,6 +699,9 @@ class RealSenseCamera(Camera):
                capture_time = time.perf_counter()

                with self.frame_lock:
+                    # Under the lock, so a late frame cannot resurrect the buffer _stop_read_thread() cleared.
+                    if stop_event.is_set():
+                        break
                    if self.use_rgb:
                        self.latest_color_frame = processed_color_frame
                    if self.use_depth:
@@ -839,4 +912,5 @@ class RealSenseCamera(Camera):
            )

        self._cleanup_resources()
+
        logger.info(f"{self} disconnected.")
@@ -102,6 +102,7 @@ class ImageServer:
                fps=self.fps,
                width=shape[1],
                height=shape[0],
+                fourcc=cfg.get("fourcc", "MJPG"),
                color_mode=ColorMode.RGB,
            )
            camera = OpenCVCamera(cam_config)
@@ -13,16 +13,41 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-from pathlib import Path
+"""Training-output persistence: checkpoints, two-phase resume, and hub publishing.

-from huggingface_hub import HfApi, snapshot_download
+Rank discipline: every function here that can
+contain a collective is documented as such and must run on ALL ranks; rank-0-only file writes
+sit under one grouped ``is_main_process()`` gate per contiguous region, placed below all
+collectives. The leaf save/load helpers carry no rank gates of their own — the exception is
+``PreTrainedPolicy._save_pretrained``, whose gate is internal because its collective gather and
+its writes live in the same method.
+"""
+
+import logging
+from importlib.resources import files
+from pathlib import Path
+from tempfile import TemporaryDirectory
+from typing import TYPE_CHECKING, Any
+
+import torch.distributed as dist
+from huggingface_hub import HfApi, ModelCard, ModelCardData, snapshot_download
 from torch.optim import Optimizer
 from torch.optim.lr_scheduler import LRScheduler

+from lerobot.__version__ import __version__
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.rewards import RewardModelConfig
 from lerobot.configs.train import TrainPipelineConfig
+from lerobot.distributed.checkpoint import (
+    is_sharded_module,
+    load_sharded_model,
+    load_sharded_optimizer,
+    save_sharded_model,
+    save_sharded_optimizer,
+)
+from lerobot.distributed.utils import is_main_process
 from lerobot.optim import (
    load_optimizer_state,
-    load_optimizer_state_dict,
    load_scheduler_state,
    save_optimizer_state,
    save_scheduler_state,
@@ -40,14 +65,39 @@ from lerobot.utils.hub import find_latest_hub_checkpoint
 from lerobot.utils.io_utils import load_json, write_json
 from lerobot.utils.random_utils import load_rng_state, save_rng_state

+if TYPE_CHECKING:
+    from accelerate import Accelerator
+
+    from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
+    from lerobot.rewards.pretrained import PreTrainedRewardModel
+

 def get_step_identifier(step: int, total_steps: int) -> str:
+    """Format a step number as the zero-padded identifier used for checkpoint directory names.
+
+    Args:
+        step (int): The training step to format.
+        total_steps (int): The total number of training steps; sets the padding width
+            (minimum 6 digits).
+
+    Returns:
+        str: The zero-padded step identifier, e.g. `"005000"`.
+    """
    num_digits = max(6, len(str(total_steps)))
    return f"{step:0{num_digits}d}"


 def get_step_checkpoint_dir(output_dir: Path, total_steps: int, step: int) -> Path:
-    """Returns the checkpoint sub-directory corresponding to the step number."""
+    """Returns the checkpoint sub-directory corresponding to the step number.
+
+    Args:
+        output_dir (Path): The training run's output directory.
+        total_steps (int): The total number of training steps; sets the identifier padding.
+        step (int): The training step of the checkpoint.
+
+    Returns:
+        Path: The checkpoint step directory, `output_dir/checkpoints/<step-identifier>`.
+    """
    step_identifier = get_step_identifier(step, total_steps)
    return output_dir / CHECKPOINTS_DIR / step_identifier

@@ -63,37 +113,15 @@ def should_save_checkpoint(step: int, save_freq: int, total_steps: int) -> bool:
    return (save_freq > 0 and step % save_freq == 0) or step == total_steps


-def save_training_step(
-    step: int, save_dir: Path, num_processes: int | None = None, batch_size: int | None = None
-) -> None:
-    state: dict = {"step": step}
-    # num_processes and batch_size are recorded so a resumed run can detect a changed world size or
-    # batch size: the sampler's resume offset is computed from the (num_processes, batch_size) that
-    # produced `step`, since both scale how many sampler positions a step consumes (see
-    # compute_sampler_state).
-    if num_processes is not None:
-        state["num_processes"] = num_processes
-    if batch_size is not None:
-        state["batch_size"] = batch_size
-    write_json(state, save_dir / TRAINING_STEP)
+def update_last_checkpoint(checkpoint_dir: Path) -> None:
+    """Point the `last` symlink in the checkpoints directory at the given checkpoint.

+    Any existing `last` symlink is replaced. The link target is relative to the checkpoints
+    directory, so the tree stays valid when the run directory is moved.

-def load_training_step(save_dir: Path) -> int:
-    training_step = load_json(save_dir / TRAINING_STEP)
-    return training_step["step"]
-
-
-def load_training_num_processes(checkpoint_dir: Path) -> int | None:
-    """World size recorded at checkpoint time, or None for checkpoints written before it was stored."""
-    return load_json(checkpoint_dir / TRAINING_STATE_DIR / TRAINING_STEP).get("num_processes")
-
-
-def load_training_batch_size(checkpoint_dir: Path) -> int | None:
-    """Per-process batch size recorded at checkpoint time, or None for older checkpoints."""
-    return load_json(checkpoint_dir / TRAINING_STATE_DIR / TRAINING_STEP).get("batch_size")
-
-
-def update_last_checkpoint(checkpoint_dir: Path) -> Path:
+    Args:
+        checkpoint_dir (Path): The checkpoint step directory the `last` link should target.
+    """
    last_checkpoint_dir = checkpoint_dir.parent / LAST_CHECKPOINT_LINK
    if last_checkpoint_dir.is_symlink():
        last_checkpoint_dir.unlink()
@@ -101,6 +129,68 @@ def update_last_checkpoint(checkpoint_dir: Path) -> Path:
    last_checkpoint_dir.symlink_to(relative_target)


+# ---------------------------------------------------------------------------------------------
+# training_step.json
+# ---------------------------------------------------------------------------------------------
+
+
+def save_training_metadata(step: int, save_dir: Path, cfg: TrainPipelineConfig) -> None:
+    """Record the step counter plus everything a resume needs to reason about topology changes.
+
+    `step` counts loop iterations (= micro-batches), so
+    the sampler resume offset is `step x batch_size x dp_world_size` with no grad-accum factor.
+    `grad_accum_steps` and the parallelism snapshot are recorded so a resume can warn precisely
+    when the optimizer-update cadence or the sharding topology changed.
+
+    Args:
+        step (int): The training step (micro-batch counter) to record.
+        save_dir (Path): The `training_state/` directory to write `training_step.json` into.
+        cfg (TrainPipelineConfig): The training config whose batch size, gradient-accumulation,
+            and parallelism settings are snapshotted alongside the step.
+    """
+    state: dict[str, Any] = {
+        "step": step,
+        "dp_world_size": cfg.parallelism.dp_world_size,
+        "batch_size": cfg.batch_size,
+        "grad_accum_steps": cfg.accelerator.gradient_accumulation.steps,
+        "parallelism": {
+            "dp_replicate": cfg.parallelism.dp_replicate,
+            "dp_shard": cfg.parallelism.dp_shard,
+            "ring_degree": cfg.parallelism.context_parallel.ring_degree,
+            "ulysses_degree": cfg.parallelism.context_parallel.ulysses_degree,
+        },
+    }
+    write_json(state, save_dir / TRAINING_STEP)
+
+
+def load_training_metadata(training_state_dir: Path) -> dict[str, Any]:
+    """Read everything `save_training_metadata` recorded, in a single pass.
+
+    Every key is always present: fields a checkpoint predates come back as None, so a caller
+    reading `metadata["batch_size"]` gets a KeyError on a typo rather than a silent None.
+
+    Args:
+        training_state_dir (Path): The checkpoint's `training_state/` directory.
+
+    Returns:
+        dict[str, Any]: `step` plus the `dp_world_size`, `batch_size`, `grad_accum_steps` and
+            `parallelism` snapshot recorded alongside it (None where not recorded).
+    """
+    state = load_json(training_state_dir / TRAINING_STEP)
+    return {
+        "step": int(state["step"]),
+        "dp_world_size": state.get("dp_world_size", state.get("num_processes")),
+        "batch_size": state.get("batch_size"),
+        "grad_accum_steps": state.get("grad_accum_steps"),
+        "parallelism": state.get("parallelism"),
+    }
+
+
+# ---------------------------------------------------------------------------------------------
+# Checkpoint save
+# ---------------------------------------------------------------------------------------------
+
+
 def save_checkpoint(
    checkpoint_dir: Path,
    step: int,
@@ -110,192 +200,301 @@ def save_checkpoint(
    scheduler: LRScheduler | None = None,
    preprocessor: PolicyProcessorPipeline | None = None,
    postprocessor: PolicyProcessorPipeline | None = None,
-    num_processes: int | None = None,
-    batch_size: int | None = None,
-    model_state_dict: dict | None = None,
-    optim_state_dict: dict | None = None,
+    accelerator: "Accelerator | None" = None,
 ) -> None:
    """This function creates the following directory structure:

    005000/  #  training step at checkpoint
    ├── pretrained_model/
    │   ├── config.json  # policy config
-    │   ├── model.safetensors  # policy weights
+    │   ├── model.safetensors  # policy weights (checkpoint_format ∈ {safetensors, safetensors_dcp}, or any non-sharded run)
+    │   ├── pytorch_model_fsdp_0/  # DCP model shards (checkpoint_format ∈ {dcp, safetensors_dcp})
    │   ├── train_config.json  # train config
-    │   ├── processor.json  # processor config (if preprocessor provided)
-    │   └── step_*.safetensors  # processor state files (if any)
+    │   ├── policy_preprocessor.json  # preprocessor config (if preprocessor provided)
+    │   ├── policy_preprocessor_step_*.safetensors  # state of the stateful preprocessor steps
+    │   ├── policy_postprocessor.json  # postprocessor config (if postprocessor provided)
+    │   └── policy_postprocessor_step_*.safetensors  # state of the stateful postprocessor steps
    └── training_state/
-        ├── optimizer_param_groups.json  #  optimizer param groups
-        ├── optimizer_state.safetensors  # optimizer state
+        ├── optimizer_param_groups.json  # optimizer param groups (non-sharded runs)
+        ├── optimizer_state.safetensors  # optimizer state (non-sharded runs)
+        ├── optimizer_0/  # DCP optimizer shards (sharded runs)
        ├── rng_state.safetensors  # rng states
-        ├── scheduler_state.json  # scheduler state
-        └── training_step.json  # training step
+        ├── scheduler_state.json  # scheduler state (if scheduler provided)
+        └── training_step.json  # training step + dp_world_size/batch_size/grad_accum + topology
+
+    Collective: MUST be called on every rank. Rank-0-only writes are gated internally, so the
+    call site needs no rank branches.

    Args:
-        cfg (TrainPipelineConfig): The training config used for this run.
+        checkpoint_dir (Path): The checkpoint step directory to write (e.g. `.../checkpoints/005000`).
        step (int): The training step at that checkpoint.
+        cfg (TrainPipelineConfig): The training config used for this run.
        policy (PreTrainedPolicy): The policy to save.
-        optimizer (Optimizer | None, optional): The optimizer to save the state from. Defaults to None.
+        optimizer (Optimizer): The optimizer to save the state from.
        scheduler (LRScheduler | None, optional): The scheduler to save the state from. Defaults to None.
-        preprocessor: The preprocessor/pipeline to save. Defaults to None.
-        postprocessor: The postprocessor/pipeline to save. Defaults to None.
-        num_processes (int | None, optional): Distributed world size to record for sample-exact
-            resume. Defaults to None (not recorded).
-        batch_size (int | None, optional): Per-process batch size to record for sample-exact
-            resume. Defaults to None (not recorded).
-        model_state_dict: Pre-gathered full (unsharded) model state dict. Required under FSDP,
-            where `policy.state_dict()` would return sharded tensors; the caller gathers it via a
-            cross-rank collective and passes it here so rank 0 can write it directly. It holds
-            FSDP's fp32 master weights and is saved as-is (the loader casts to the policy dtype on
-            read). When None (DDP / single-GPU), the model is saved the normal way. Defaults to None.
-        optim_state_dict: Pre-gathered full (unsharded) optimizer state dict. Required under FSDP
-            (gathered alongside `model_state_dict` via `gather_fsdp_state_dicts`); saved in the same
-            safetensors format as the single-GPU path. When None, `optimizer.state_dict()` is used.
+        preprocessor (PolicyProcessorPipeline | None, optional): The preprocessor/pipeline to save.
            Defaults to None.
+        postprocessor (PolicyProcessorPipeline | None, optional): The postprocessor/pipeline to save.
+            Defaults to None.
+        accelerator (Accelerator | None, optional): The accelerator the policy was prepared with;
+            used to unwrap the model and required on sharded runs, where it owns the DCP save
+            channels. Defaults to None (plain single-process saves).
    """
    pretrained_dir = checkpoint_dir / PRETRAINED_MODEL_DIR
-    policy.save_pretrained(pretrained_dir, state_dict=model_state_dict)
-    cfg.save_pretrained(pretrained_dir)
+    fmt = cfg.checkpoint_format
+    policy_to_save = accelerator.unwrap_model(policy) if accelerator is not None else policy
+    sharded = is_sharded_module(policy_to_save)
+
+    # -- model artifact(s): the two collective-capable calls ----------------------------------
    if cfg.peft is not None:
-        # When using PEFT, policy.save_pretrained will only write the adapter weights + config, not the
-        # policy config which we need for loading the model. In this case we'll write it ourselves.
-        policy.config.save_pretrained(pretrained_dir)
-    if preprocessor is not None:
-        preprocessor.save_pretrained(pretrained_dir)
-    if postprocessor is not None:
-        postprocessor.save_pretrained(pretrained_dir)
+        # PeftModel.save_pretrained is an external API with no internal rank gate, and the
+        # adapters are replicated (PEFT x sharded is rejected at validation): main rank writes.
+        if is_main_process():
+            policy_to_save.save_pretrained(pretrained_dir)
+    elif fmt.wants_safetensors or not sharded:
+        # Collective when sharded (full gather); writes happen on the main process only in all
+        # multi-rank layouts (the gate lives inside _save_pretrained, next to its collective gather).
+        policy_to_save.save_pretrained(pretrained_dir)
+    if fmt.wants_dcp and sharded:
+        save_sharded_model(accelerator, policy_to_save, pretrained_dir)
+
+    # -- sidecar configs: ONE gate for the whole contiguous rank-0-only region ----------------
+    if is_main_process():
+        if fmt.wants_dcp and not fmt.wants_safetensors:
+            # save_pretrained did not run: keep the DCP-only checkpoint self-describing.
+            policy_to_save.config.save_pretrained(pretrained_dir)
+        cfg.save_pretrained(pretrained_dir)
+        if cfg.peft is not None:
+            # PEFT's save_pretrained writes only adapter weights + config; the policy config
+            # needed to reload the base model is written explicitly.
+            policy_to_save.config.save_pretrained(pretrained_dir)
+        if preprocessor is not None:
+            preprocessor.save_pretrained(pretrained_dir)
+        if postprocessor is not None:
+            postprocessor.save_pretrained(pretrained_dir)
+
    save_training_state(
-        checkpoint_dir,
-        step,
-        optimizer,
-        scheduler,
-        num_processes=num_processes,
-        batch_size=batch_size,
-        optim_state_dict=optim_state_dict,
+        checkpoint_dir, step, cfg, optimizer, scheduler, accelerator, sharded=sharded, model=policy_to_save
    )
+    if accelerator is not None:
+        accelerator.wait_for_everyone()


 def save_training_state(
    checkpoint_dir: Path,
-    train_step: int,
-    optimizer: Optimizer | None = None,
+    step: int,
+    cfg: TrainPipelineConfig,
+    optimizer: Optimizer | dict[str, Optimizer] | None = None,
    scheduler: LRScheduler | None = None,
-    num_processes: int | None = None,
-    batch_size: int | None = None,
-    optim_state_dict: dict | None = None,
+    accelerator: "Accelerator | None" = None,
+    *,
+    sharded: bool = False,
+    model: PreTrainedPolicy | None = None,
 ) -> None:
-    """
-    Saves the training step, optimizer state, scheduler state, and rng state.
+    """Write training_state/. Collective under sharding: call on every rank.

    Args:
-        save_dir (Path): The directory to save artifacts to.
-        train_step (int): Current training step.
-        optimizer (Optimizer | None, optional): The optimizer from which to save the state_dict.
+        checkpoint_dir (Path): The checkpoint step directory; `training_state/` is created inside it.
+        step (int): The training step at that checkpoint.
+        cfg (TrainPipelineConfig): The training config used for this run (its topology and
+            accumulation settings are recorded in `training_step.json`).
+        optimizer (Optimizer | dict[str, Optimizer] | None, optional): The optimizer(s) to save
+            the state from. Defaults to None.
+        scheduler (LRScheduler | None, optional): The scheduler to save the state from.
            Defaults to None.
-        scheduler (LRScheduler | None, optional): The scheduler from which to save the state_dict.
-            Defaults to None.
-        num_processes (int | None, optional): Distributed world size to record. Defaults to None.
-        batch_size (int | None, optional): Per-process batch size to record. Defaults to None.
-        optim_state_dict: Pre-gathered full optimizer state dict (for FSDP). Saved instead of
-            `optimizer.state_dict()` when provided. Defaults to None.
+        accelerator (Accelerator | None, optional): Required when `sharded` is True — it owns
+            the DCP optimizer save channel. Defaults to None.
+        sharded (bool): The model's sharding state, computed once in `save_checkpoint` and
+            threaded here so the two sites cannot disagree. Defaults to False.
+        model (PreTrainedPolicy | None, optional): Required only for the sharded optimizer
+            channel: torch's optimizer DCP APIs are model-coupled (the state dict is keyed by
+            model FQNs), so accelerate's `save_fsdp_optimizer` needs the sharded module
+            alongside the optimizer. Defaults to None.
    """
    save_dir = checkpoint_dir / TRAINING_STATE_DIR
+    # All ranks: the directory must exist before the DCP optimizer collective writes into it
+    # (exist_ok makes the concurrent mkdir race-free on shared filesystems).
    save_dir.mkdir(parents=True, exist_ok=True)
-    save_training_step(train_step, save_dir, num_processes=num_processes, batch_size=batch_size)
-    save_rng_state(save_dir)
-    if optimizer is not None:
-        save_optimizer_state(optimizer, save_dir, optim_state_dict=optim_state_dict)
-    if scheduler is not None:
-        save_scheduler_state(scheduler, save_dir)
+
+    if optimizer is not None and sharded:
+        if accelerator is None or model is None:
+            raise ValueError("Saving a sharded optimizer state requires the accelerator and model.")
+        # Collective — all ranks write their DCP shards into optimizer_0/.
+        save_sharded_optimizer(accelerator, optimizer, model, save_dir)
+
+    if is_main_process():  # ONE grouped gate for the whole rank-0-only region
+        save_training_metadata(step, save_dir, cfg)
+        save_rng_state(save_dir)
+        if scheduler is not None:
+            save_scheduler_state(scheduler, save_dir)
+        if optimizer is not None and not sharded:
+            save_optimizer_state(optimizer, save_dir)


-def load_training_state(
-    checkpoint_dir: Path, optimizer: Optimizer, scheduler: LRScheduler | None, load_optimizer: bool = True
-) -> tuple[int, Optimizer, LRScheduler | None]:
-    """
-    Loads the training step, optimizer state, scheduler state, and rng state.
-    This is used to resume a training run.
+# ---------------------------------------------------------------------------------------------
+# Two-phase resume
+# ---------------------------------------------------------------------------------------------
+
+
+def resume_before_prepare(cfg: TrainPipelineConfig) -> int:
+    """Phase 1 — before `accelerator.prepare()`: restore RNG and return the step counter.
+
+    Pure loaders only. The sampler resume offset is *derived* from the returned step inside the
+    dataloader factory, and everything bound to sharded objects (model DCP shards, optimizer,
+    scheduler) loads in `resume_after_prepare`.

    Args:
-        checkpoint_dir (Path): The checkpoint directory. Should contain a 'training_state' dir.
-        optimizer (Optimizer): The optimizer to load the state_dict to.
-        scheduler (LRScheduler | None): The scheduler to load the state_dict to (can be None).
-        load_optimizer (bool, optional): Whether to load the optimizer state from disk. Defaults to
-            True. Set to False under FSDP, where the sharded optimizer state must be loaded after
-            `accelerator.prepare()` via `load_fsdp_optimizer_state` (the optimizer is returned
-            untouched here).
+        cfg (TrainPipelineConfig): The resumed training config; `cfg.checkpoint_path` locates
+            the checkpoint to restore from.
+
+    Returns:
+        int: The training step recorded in the checkpoint (micro-batch counter).

    Raises:
-        NotADirectoryError: If 'checkpoint_dir' doesn't contain a 'training_state' dir
-
-    Returns:
-        tuple[int, Optimizer, LRScheduler | None]: training step, optimizer and scheduler with their
-            state_dict loaded.
+        NotADirectoryError: If the checkpoint has no `training_state/` directory.
+        ValueError: If the resumed topology crosses the sharded/non-sharded boundary relative
+            to the one recorded in the checkpoint.
    """
-    training_state_dir = checkpoint_dir / TRAINING_STATE_DIR
+    training_state_dir = cfg.checkpoint_path / TRAINING_STATE_DIR
    if not training_state_dir.is_dir():
        raise NotADirectoryError(training_state_dir)
-
+    metadata = load_training_metadata(training_state_dir)
+    _guard_resume_changes(cfg, metadata)
    load_rng_state(training_state_dir)
-    step = load_training_step(training_state_dir)
-    if load_optimizer:
-        optimizer = load_optimizer_state(optimizer, training_state_dir)
+    return metadata["step"]
+
+
+def _guard_resume_changes(cfg: TrainPipelineConfig, metadata: dict[str, Any]) -> None:
+    """Check the resumed run settings against the ones recorded in the checkpoint.
+
+    Two tiers, both driven by the checkpoint's recorded parallelism snapshot:
+
+    - **Hard error** when the resume crosses the sharded/non-sharded boundary in either
+      direction: the checkpoint's training-state artifacts only support resuming on the same
+      kind of topology (resharding works across sizes, not across kinds). Checkpoints without
+      a recorded snapshot skip this check.
+    - **One warning** naming every other recorded setting that differs — those changes are
+      legal (DCP reshards weights and optimizer state across topologies and the sampler offset
+      adapts), but a changed ``grad_accum_steps`` shifts the optimizer-update cadence, so the
+      resume says precisely what differs. The sampler-exactness warnings
+      (``dp_world_size``/``batch_size``) live with the sampler math in the dataloader factory.
+
+    Args:
+        cfg (TrainPipelineConfig): The resumed training config, compared against the settings
+            recorded in the checkpoint.
+        metadata (dict[str, Any]): The checkpoint's recorded training metadata, as returned by
+            `load_training_metadata`.
+
+    Raises:
+        ValueError: If the checkpoint records a sharded topology and the resumed run is
+            non-sharded, or vice versa.
+    """
+    snapshot = metadata["parallelism"]
+
+    if snapshot is not None:
+        recorded_sharded = (
+            snapshot.get("dp_shard", 1) != 1
+            or snapshot.get("ring_degree", 1) * snapshot.get("ulysses_degree", 1) > 1
+        )
+        if recorded_sharded != cfg.parallelism.is_sharded:
+            raise ValueError(
+                f"Cannot resume: the checkpoint was written with a "
+                f"{'sharded' if recorded_sharded else 'non-sharded'} topology "
+                f"(dp_replicate={snapshot.get('dp_replicate')}, dp_shard={snapshot.get('dp_shard')}) "
+                f"but this run is {'sharded' if cfg.parallelism.is_sharded else 'non-sharded'} "
+                f"(dp_replicate={cfg.parallelism.dp_replicate}, dp_shard={cfg.parallelism.dp_shard})."
+            )
+
+    recorded = {
+        "grad_accum_steps": (
+            metadata["grad_accum_steps"],
+            cfg.accelerator.gradient_accumulation.steps,
+        ),
+    }
+    if snapshot is not None:
+        recorded.update(
+            {
+                "dp_replicate": (snapshot.get("dp_replicate"), cfg.parallelism.dp_replicate),
+                "dp_shard": (snapshot.get("dp_shard"), cfg.parallelism.dp_shard),
+                "ring_degree": (
+                    snapshot.get("ring_degree"),
+                    cfg.parallelism.context_parallel.ring_degree,
+                ),
+                "ulysses_degree": (
+                    snapshot.get("ulysses_degree"),
+                    cfg.parallelism.context_parallel.ulysses_degree,
+                ),
+            }
+        )
+    changed = [f"{key}: {was} -> {now}" for key, (was, now) in recorded.items() if was not in (None, now)]
+    if changed and is_main_process():
+        logging.warning(
+            "Resuming with settings that differ from the checkpoint: " + "; ".join(changed) + ". "
+            "Topology changes reshard safely via DCP; a changed grad_accum_steps shifts the "
+            "optimizer-update cadence (the step counter keeps counting micro-batches)."
+        )
+
+
+def resume_after_prepare(
+    cfg: TrainPipelineConfig,
+    accelerator: "Accelerator",
+    policy: PreTrainedPolicy,
+    optimizer: Optimizer | dict[str, Optimizer],
+    scheduler: LRScheduler | None,
+) -> None:
+    """Phase 2 — after `accelerator.prepare()`: model (DCP) -> optimizer -> scheduler.
+
+    Collective under sharding: call on every rank. The model-weight source follows the
+    checkpoint's own recorded `checkpoint_format` (on resume, `cfg` was parsed from the
+    checkpoint's train_config.json): DCP-bearing formats load shards here into the prepared
+    model (whose construction skipped the safetensors load); the safetensors format was already
+    loaded by `from_pretrained` before sharding — no model step here.
+
+    Args:
+        cfg (TrainPipelineConfig): The resumed training config; `cfg.checkpoint_path` locates
+            the checkpoint and `cfg.checkpoint_format` selects the model-weight source.
+        accelerator (Accelerator): The accelerator the policy was prepared with; it unwraps the
+            model and owns the DCP load channels.
+        policy (PreTrainedPolicy): The prepared (possibly sharded) policy to load weights into.
+        optimizer (Optimizer | dict[str, Optimizer]): The prepared optimizer(s) to restore.
+        scheduler (LRScheduler | None): The scheduler to restore, or None if the run has none.
+
+    Raises:
+        FileNotFoundError: If the checkpoint format declares DCP model shards but the shard
+            directory is missing (e.g. it was pruned before upload).
+    """
+    checkpoint_dir = cfg.checkpoint_path
+    pretrained_dir = checkpoint_dir / PRETRAINED_MODEL_DIR
+    training_state_dir = checkpoint_dir / TRAINING_STATE_DIR
+    unwrapped = accelerator.unwrap_model(policy)
+    sharded = is_sharded_module(unwrapped)
+
+    if cfg.checkpoint_format.wants_dcp:
+        from accelerate.utils.constants import FSDP_MODEL_NAME
+
+        dcp_dir = pretrained_dir / f"{FSDP_MODEL_NAME}_0"
+        if not dcp_dir.is_dir():
+            raise FileNotFoundError(
+                f"checkpoint_format={cfg.checkpoint_format.value} declares DCP model shards, "
+                f"but {dcp_dir} is missing. If the shards were pruned, convert what remains "
+                "with `lerobot-convert-dcp` or resume from a safetensors checkpoint."
+            )
+        load_sharded_model(accelerator, unwrapped, pretrained_dir)
+
+    if sharded:
+        # Requires the prepared optimizer: FSDP2's prepare rebinds param groups to DTensors but
+        # never migrates optimizer.state — DCP reshards it here (works across topology changes).
+        load_sharded_optimizer(accelerator, optimizer, unwrapped, training_state_dir)
+    else:
+        load_optimizer_state(optimizer, training_state_dir)
+
    if scheduler is not None:
-        scheduler = load_scheduler_state(scheduler, training_state_dir)
-
-    return step, optimizer, scheduler
+        load_scheduler_state(scheduler, training_state_dir)


-def gather_fsdp_state_dicts(model, optimizer) -> tuple[dict, dict]:
-    """Gather the full (unsharded) model and optimizer state dicts under FSDP.
-
-    `model.state_dict()` and `FSDP.optim_state_dict(...)` are cross-rank collectives, so this must be
-    called on *every* rank with the prepared (FSDP-wrapped) `model` and `optimizer`. With
-    `rank0_only=True` and `offload_to_cpu=True`, every rank runs the all-gather but only rank 0
-    materializes the full dicts (the others get empty dicts) and they are kept on CPU to bound GPU
-    memory. The returned optimizer state dict is keyed by parameter FQNs and is world-size
-    independent; `load_fsdp_optimizer_state` reshards it on resume.
-
-    Returns:
-        (model_state_dict, optim_state_dict): full dicts on rank 0, empty dicts on other ranks.
-    """
-    from torch.distributed.fsdp import (
-        FullOptimStateDictConfig,
-        FullStateDictConfig,
-        FullyShardedDataParallel as FSDP,  # noqa F401
-        StateDictType,
-    )
-
-    state_cfg = FullStateDictConfig(offload_to_cpu=True, rank0_only=True)
-    optim_cfg = FullOptimStateDictConfig(offload_to_cpu=True, rank0_only=True)
-    with FSDP.state_dict_type(model, StateDictType.FULL_STATE_DICT, state_cfg, optim_cfg):
-        model_state_dict = model.state_dict()
-        optim_state_dict = FSDP.optim_state_dict(model, optimizer)
-    return model_state_dict, optim_state_dict
-
-
-def load_fsdp_optimizer_state(model, optimizer, checkpoint_dir: Path) -> None:
-    """Load the FSDP optimizer state (saved as safetensors) and reshard it into the optimizer.
-
-    This is a cross-rank collective and must be called on every rank *after* `accelerator.prepare()`
-    with the prepared (FSDP-wrapped) `model` and `optimizer`. The saved state is the full,
-    world-size-independent optimizer state (keyed by parameter FQNs); `FSDP.optim_state_dict_to_load`
-    reshards it to the current FSDP topology, so resume on a different number of GPUs works.
-    """
-    from torch.distributed.fsdp import (
-        FullOptimStateDictConfig,
-        FullStateDictConfig,
-        FullyShardedDataParallel as FSDP,  # noqa F401
-        StateDictType,
-    )
-
-    # Every rank reads the same full state from the (shared) checkpoint dir, so rank0_only=False.
-    full_osd = load_optimizer_state_dict(checkpoint_dir / TRAINING_STATE_DIR)
-    state_cfg = FullStateDictConfig(rank0_only=False)
-    optim_cfg = FullOptimStateDictConfig(rank0_only=False)
-    with FSDP.state_dict_type(model, StateDictType.FULL_STATE_DICT, state_cfg, optim_cfg):
-        sharded_osd = FSDP.optim_state_dict_to_load(model=model, optim=optimizer, optim_state_dict=full_osd)
-    optimizer.load_state_dict(sharded_osd)
+# ---------------------------------------------------------------------------------------------
+# Hub: checkpoint push (resume artifact) and publishing (distribution artifact)
+# ---------------------------------------------------------------------------------------------


 def push_checkpoint_to_hub(
@@ -311,6 +510,16 @@ def push_checkpoint_to_hub(
    The model repo is created idempotently, and the commit is tagged with the
    checkpoint step so a checkpoint can be recovered with
    --policy.pretrained_revision=<step> instead of a commit sha.
+
+    The directory is uploaded verbatim — including DCP shards under the DCP formats: this tree
+    exists for *resume*, not distribution, and `resolve_resume_checkpoint` downloads it back
+    symmetrically.
+
+    Args:
+        checkpoint_dir (Path): The local checkpoint step directory to upload.
+        repo_id (str): The Hub model repo to push to (created idempotently if missing).
+        private (bool | None): Whether a newly created repo should be private. Defaults to
+            None (public unless the organization's default is private).
    """
    api = HfApi()
    api.create_repo(repo_id=repo_id, repo_type="model", private=private, exist_ok=True)
@@ -338,6 +547,16 @@ def resolve_resume_checkpoint(repo_id: str, output_dir: Path) -> Path:
    into `output_dir/checkpoints/<step>/`, recreate the local `last` symlink, and return that local
    checkpoint dir. Used to resume training from the Hub on a machine (or HF Jobs pod) that does not
    have the original local run dir.
+
+    Args:
+        repo_id (str): The Hub model repo holding `checkpoints/<step>/` subtrees.
+        output_dir (Path): The local run directory to download the checkpoint into.
+
+    Returns:
+        Path: The local checkpoint step directory, `output_dir/checkpoints/<step>`.
+
+    Raises:
+        FileNotFoundError: If the repo contains no checkpoints under `checkpoints/`.
    """
    latest = find_latest_hub_checkpoint(repo_id)
    if latest is None:
@@ -354,3 +573,214 @@ def resolve_resume_checkpoint(repo_id: str, output_dir: Path) -> Path:
    checkpoint_dir = output_dir / latest
    update_last_checkpoint(checkpoint_dir)
    return checkpoint_dir
+
+
+def publish_trained_model(
+    cfg: TrainPipelineConfig,
+    model: "PreTrainedPolicy | PreTrainedRewardModel",
+    preprocessor: PolicyProcessorPipeline | None,
+    postprocessor: PolicyProcessorPipeline | None,
+    dataset_meta: "LeRobotDatasetMetadata | None",
+    *,
+    peft_model: Any | None = None,
+) -> None:
+    """Publish the complete training bundle as a distributable model repo.
+
+    Collective-safe: call on ALL ranks — the model commit gathers sharded weights through
+    `save_pretrained`; uploads happen on the main process only (gated inside
+    `HubMixin.push_to_hub` and here). Commits, in order: (1) the model (skipped for PEFT —
+    adapters replace full weights), (2) the preprocessor, (3) the postprocessor, (4) the bundle
+    sidecar: README.md model card + train_config.json (+ adapter weights and the wrapped
+    policy's config in the PEFT case). Every commit uploads a freshly assembled directory, so
+    a published repo carries only the distributable artifacts.
+
+    Args:
+        cfg (TrainPipelineConfig): The training config; saved as `train_config.json` and used
+            to render the model card.
+        model (PreTrainedPolicy | PreTrainedRewardModel): The trained model to publish; its
+            config supplies the target repo id, visibility, license, and tags.
+        preprocessor (PolicyProcessorPipeline | None): The preprocessor pipeline to publish
+            alongside the model, if any.
+        postprocessor (PolicyProcessorPipeline | None): The postprocessor pipeline to publish
+            alongside the model, if any.
+        dataset_meta (LeRobotDatasetMetadata | None): Dataset metadata for the model card, if
+            available.
+        peft_model (Any | None): The PEFT wrapper when training adapters; its adapter weights
+            replace the full model weights in the published repo. Defaults to None.
+
+    Raises:
+        ValueError: If the model config carries no repo id (`--policy.repo_id`).
+    """
+    model_cfg = model.config
+    repo_id = model_cfg.repo_id
+    if not repo_id:
+        raise ValueError("Publishing requires a repo id (--policy.repo_id).")
+    ignore = ["*.tmp", "*.log"]
+
+    if peft_model is None:
+        # Calls are made on the exact objects that own each method (never through PEFT's
+        # attribute forwarding), so the peft branch below never touches this path.
+        model.push_to_hub(repo_id, private=model_cfg.private, ignore_patterns=ignore)
+    if preprocessor is not None:
+        preprocessor.push_to_hub(repo_id, private=model_cfg.private)
+    if postprocessor is not None:
+        postprocessor.push_to_hub(repo_id, private=model_cfg.private)
+
+    if is_main_process():
+        api = HfApi()
+        repo_id = api.create_repo(repo_id=repo_id, private=model_cfg.private, exist_ok=True).repo_id
+        with TemporaryDirectory(ignore_cleanup_errors=True) as tmp:
+            saved_path = Path(tmp) / repo_id
+            saved_path.mkdir(parents=True, exist_ok=True)
+            if peft_model is not None:
+                peft_model.save_pretrained(saved_path)  # adapter weights + adapter config
+                model.config.save_pretrained(saved_path)  # PEFT cannot write the policy config
+            card = generate_model_card(model_cfg, cfg=cfg, dataset_meta=dataset_meta)
+            card.save(str(saved_path / "README.md"))
+            cfg.save_pretrained(saved_path)  # train_config.json
+            commit_info = api.upload_folder(
+                repo_id=repo_id,
+                repo_type="model",
+                folder_path=saved_path,
+                commit_message="Upload model card and train config",
+                allow_patterns=["*.safetensors", "*.json", "*.yaml", "*.md"],
+                ignore_patterns=ignore,
+            )
+        # Contract: lerobot.jobs.hf.submit_to_hf watches for this exact "Model pushed to <url>"
+        # line to end a remote run early. Keep the wording and URL format in sync.
+        logging.info(f"Model pushed to {commit_info.repo_url.url}")
+
+    if dist.is_initialized():
+        dist.barrier()
+
+
+# ---------------------------------------------------------------------------------------------
+# Model card
+# ---------------------------------------------------------------------------------------------
+
+_BASE_MODEL_MAPPING = {
+    "smolvla": "lerobot/smolvla_base",
+    "pi0": "lerobot/pi0_base",
+    "pi05": "lerobot/pi05_base",
+    "pi0_fast": "lerobot/pi0fast-base",
+    "xvla": "lerobot/xvla-base",
+}
+
+
+def build_card_context(
+    cfg: TrainPipelineConfig | None,
+    dataset_meta: "LeRobotDatasetMetadata | None",
+    input_features: dict | None,
+    output_features: dict | None,
+) -> dict:
+    """Collect optional data for the model-card template.
+
+    Returns plain values only (no Markdown) — the template in
+    ``lerobot/templates/lerobot_modelcard_template.md`` decides how and whether to show
+    each one. Everything is best-effort: anything unavailable is left empty/None and the
+    template simply skips that section, so this never breaks a Hub push.
+
+    Args:
+        cfg (TrainPipelineConfig | None): The training config supplying the training section,
+            if available.
+        dataset_meta (LeRobotDatasetMetadata | None): Dataset metadata supplying the dataset,
+            robot-type, and camera sections, if available.
+        input_features (dict | None): The policy's input feature declarations, if any.
+        output_features (dict | None): The policy's output feature declarations, if any.
+
+    Returns:
+        dict: Template context with `training`, `input_features`, `output_features`,
+            `dataset`, `robot_type`, and `cameras` entries; unavailable pieces stay
+            empty/None.
+    """
+    context = {
+        "training": None,
+        "input_features": input_features or {},
+        "output_features": output_features or {},
+        "dataset": None,
+        "robot_type": None,
+        "cameras": [],
+    }
+
+    if cfg is not None:
+        optimizer = getattr(cfg, "optimizer", None)
+        context["training"] = {
+            "steps": cfg.steps,
+            "batch_size": cfg.batch_size,
+            "seed": cfg.seed,
+            "optimizer": getattr(optimizer, "type", None) if optimizer else None,
+            "lr": getattr(optimizer, "lr", None) if optimizer else None,
+            "lerobot_version": __version__,
+        }
+
+    if dataset_meta is not None:
+        context["dataset"] = {
+            "repo_id": dataset_meta.repo_id,
+            "episodes": dataset_meta.total_episodes,
+            "frames": dataset_meta.total_frames,
+            "fps": dataset_meta.fps,
+            "tasks": [str(task) for task in dataset_meta.tasks.index],
+        }
+        context["robot_type"] = dataset_meta.robot_type
+        context["cameras"] = [key.split(".")[-1] for key in dataset_meta.camera_keys]
+
+    return context
+
+
+def generate_model_card(
+    model_cfg: PreTrainedConfig | RewardModelConfig,
+    cfg: TrainPipelineConfig | None = None,
+    dataset_meta: "LeRobotDatasetMetadata | None" = None,
+) -> ModelCard:
+    """Render the LeRobot model card for a trained policy or reward model.
+
+    A free function on purpose: every template variable comes from arguments — the model
+    config, the training config, and the dataset metadata — none from a live model, so a card
+    can also be rendered from a checkpoint's `config.json` alone (see `lerobot-convert-dcp`).
+    The config type selects the template: reward models get the reward-model card, policies the
+    policy card with the training/dataset sections.
+
+    Args:
+        model_cfg (PreTrainedConfig | RewardModelConfig): The model config providing type,
+            license, tags, repo id, and — for policies — the feature declarations.
+        cfg (TrainPipelineConfig | None, optional): The training config for the training and
+            dataset card sections. Defaults to None.
+        dataset_meta (LeRobotDatasetMetadata | None, optional): Dataset metadata for the
+            dataset card sections. Defaults to None.
+
+    Returns:
+        ModelCard: The rendered and validated LeRobot model card.
+    """
+    model_type = model_cfg.type
+    base_model = _BASE_MODEL_MAPPING.get(model_type)
+
+    if isinstance(model_cfg, RewardModelConfig):
+        tags = {"robotics", "lerobot", "reward-model", model_type}
+        template_card = (
+            files("lerobot.templates")
+            .joinpath("lerobot_rewardmodel_modelcard_template.md")
+            .read_text("utf-8")
+        )
+        context: dict[str, Any] = {}  # the reward template renders from card_data alone
+    else:
+        tags = {"robotics", "lerobot", model_type}
+        template_card = (
+            files("lerobot.templates").joinpath("lerobot_modelcard_template.md").read_text("utf-8")
+        )
+        context = build_card_context(cfg, dataset_meta, model_cfg.input_features, model_cfg.output_features)
+        # Used by the template to pre-fill commands and the "Fine-tuned from" line.
+        context["policy_repo_id"] = model_cfg.repo_id
+        context["base_model"] = base_model
+
+    card_data = ModelCardData(
+        license=model_cfg.license or "apache-2.0",
+        library_name="lerobot",
+        pipeline_tag="robotics",
+        tags=list(tags.union(model_cfg.tags or [])),
+        model_name=model_type,
+        datasets=cfg.dataset.repo_id if cfg is not None else None,
+        base_model=base_model,
+    )
+    card = ModelCard.from_template(card_data, template_str=template_card, **context)
+    card.validate()
+    return card
@@ -22,7 +22,7 @@ Import them directly: ``from lerobot.configs.train import TrainPipelineConfig``
 """

 from .dataset import DatasetRecordConfig
-from .default import DatasetConfig, EvalConfig, JobConfig, PeftConfig, WandBConfig
+from .default import DatasetConfig, EMAConfig, EvalConfig, JobConfig, PeftConfig, WandBConfig
 from .policies import PreTrainedConfig
 from .recipe import MessageTurn, TrainingRecipe, load_recipe
 from .types import (
@@ -57,6 +57,7 @@ __all__ = [
    # Config classes
    "DatasetRecordConfig",
    "DatasetConfig",
+    "EMAConfig",
    "EvalConfig",
    "JobConfig",
    "MessageTurn",
@@ -0,0 +1,273 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Execution-runtime configuration: everything handed to (or applied by) the `Accelerator`.
+
+Each sub-config mirrors the plain-typed subset of the corresponding accelerate object and
+builds it at runtime (the way ``OptimizerConfig.build()`` constructs a ``torch.optim.Optimizer``),
+so the whole tree round-trips through the CLI and ``train_config.json`` and parsing a config
+never imports accelerate.
+"""
+
+from dataclasses import dataclass, field
+from enum import Enum
+from typing import TYPE_CHECKING
+
+from lerobot.configs.parallelism import ParallelismConfig
+
+if TYPE_CHECKING:
+    from accelerate import Accelerator
+    from accelerate.utils import (
+        DistributedDataParallelKwargs,
+        FullyShardedDataParallelPlugin,
+        GradientAccumulationPlugin,
+    )
+
+
+@dataclass
+class FSDPConfig:
+    """Mirror of the `FullyShardedDataParallelPlugin` subset LeRobot supports (FSDP2 only).
+
+    Exactly one wrap policy applies: `wrap_modules` (module *class names* forming the FSDP
+    units — and, later, the activation-checkpointing units) or `min_num_params` (size-based).
+    When both are None, the policy's own `_fsdp_wrap_modules` declaration is used; a run where
+    no wrap source exists at all fails loudly rather than silently wrapping only the root.
+    """
+
+    reshard_after_forward: bool = True
+    wrap_modules: list[str] | None = None
+    min_num_params: int | None = None
+    cpu_offload: bool = False
+    # Regex matched against module FQNs to exclude their parameters from sharding.
+    ignored_modules: str | None = None
+
+    def __post_init__(self) -> None:
+        """Validate the wrap-policy fields.
+
+        Raises:
+            ValueError: If both ``wrap_modules`` and ``min_num_params`` are set (they are
+                mutually exclusive wrap policies), or if ``min_num_params`` is < 1.
+        """
+        if self.wrap_modules is not None and self.min_num_params is not None:
+            raise ValueError(
+                "fsdp.wrap_modules and fsdp.min_num_params are mutually exclusive wrap policies."
+            )
+        if self.min_num_params is not None and self.min_num_params < 1:
+            raise ValueError(f"fsdp.min_num_params must be >= 1, got {self.min_num_params}.")
+
+    def build_plugin(self) -> "FullyShardedDataParallelPlugin":
+        """Build the FSDP2 plugin for `Accelerator(fsdp_plugin=...)`.
+
+        Returns:
+            FullyShardedDataParallelPlugin: FSDP2 (`fsdp_version=2`) plugin carrying the
+                mirrored wrap policy, resharding, CPU-offload, and ignored-modules settings.
+        """
+        from accelerate.utils import FullyShardedDataParallelPlugin
+
+        use_size_policy = self.min_num_params is not None
+        return FullyShardedDataParallelPlugin(
+            fsdp_version=2,
+            reshard_after_forward=self.reshard_after_forward,
+            auto_wrap_policy="size_based_wrap" if use_size_policy else "transformer_based_wrap",
+            # May legitimately still be None here: the policy-declared default is applied right
+            # before `accelerator.prepare()` (see lerobot.distributed.factory.set_fsdp_wrap_modules).
+            transformer_cls_names_to_wrap=list(self.wrap_modules) if self.wrap_modules else None,
+            min_num_params=self.min_num_params,
+            cpu_offload=self.cpu_offload,
+            ignored_modules=self.ignored_modules,
+            # state_dict_type stays at the FSDP2 default (SHARDED_STATE_DICT) and is never
+            # switched: full gathers go through torch's state-dict API, which does not consult
+            # the plugin. activation_checkpointing stays False: AC is LeRobot-owned.
+        )
+
+
+@dataclass
+class DDPConfig:
+    """Mirror of the `DistributedDataParallelKwargs` subset LeRobot exposes."""
+
+    # Today's in-script default, kept for models with conditional computation.
+    find_unused_parameters: bool = True
+    gradient_as_bucket_view: bool = False
+    static_graph: bool = False
+
+    def build_kwargs_handler(self) -> "DistributedDataParallelKwargs":
+        """Build the DDP kwargs handler for `Accelerator(kwargs_handlers=[...])`.
+
+        Returns:
+            DistributedDataParallelKwargs: Handler carrying the mirrored DDP fields, applied
+                by accelerate when it wraps the model in `DistributedDataParallel`.
+        """
+        from accelerate.utils import DistributedDataParallelKwargs
+
+        return DistributedDataParallelKwargs(
+            find_unused_parameters=self.find_unused_parameters,
+            gradient_as_bucket_view=self.gradient_as_bucket_view,
+            static_graph=self.static_graph,
+        )
+
+
+@dataclass
+class GradientAccumulationConfig:
+    """Mirror of the `GradientAccumulationPlugin` subset LeRobot supports.
+
+    Only the step count is a knob. ``sync_with_dataloader`` is pinned to False by
+    :meth:`build_plugin`: the training loop cycles a finite dataloader, so accelerate's default
+    of syncing at every dataloader end would force an optimizer step at every dataset epoch
+    boundary instead of every ``steps`` micro-batches.
+    """
+
+    steps: int = 1
+
+    def __post_init__(self) -> None:
+        """Validate the accumulation step count.
+
+        Raises:
+            ValueError: If ``steps`` is < 1.
+        """
+        if self.steps < 1:
+            raise ValueError(f"gradient_accumulation.steps must be >= 1, got {self.steps}.")
+
+    def build_plugin(self) -> "GradientAccumulationPlugin":
+        """Build the plugin for `Accelerator(gradient_accumulation_plugin=...)`.
+
+        A named plugin argument, not a `kwargs_handlers` entry: accelerate consumes this object
+        through its dedicated constructor parameter — the `KwargsHandler` base class only lends
+        it `to_kwargs()`, so the consumption site, not the inheritance, decides its role.
+
+        Returns:
+            GradientAccumulationPlugin: Carrying the mirrored step count, with
+                ``sync_with_dataloader=False`` pinned (see the class docstring).
+        """
+        from accelerate.utils import GradientAccumulationPlugin
+
+        return GradientAccumulationPlugin(num_steps=self.steps, sync_with_dataloader=False)
+
+
+@dataclass
+class CompileConfig:
+    """torch.compile knobs — a configured placeholder: wiring lands in a later round.
+
+    The setup-order contract it will follow is already fixed: compile applies
+    after CP dispatch install and activation checkpointing, before `fully_shard`, regionally
+    (per wrap unit) — the only combination proven with FSDP2.
+    """
+
+    enabled: bool = False
+    backend: str = "inductor"
+    mode: str | None = None
+    regional: bool = True
+
+
+class ActivationCheckpointingMode(str, Enum):
+    NONE = "none"
+    FULL = "full"
+
+
+@dataclass
+class ActivationCheckpointingConfig:
+    """Activation-checkpointing knobs — a configured placeholder: wiring lands in a later round.
+
+    AC units will coincide with the FSDP wrap units (one declaration drives both), applied
+    before torch.compile and `fully_shard` (the same ordering contract as CompileConfig).
+    """
+
+    mode: ActivationCheckpointingMode = ActivationCheckpointingMode.NONE
+
+
+@dataclass
+class AcceleratorConfig:
+    """Builds the `Accelerator` — the runtime counterpart of the `parallelism` topology.
+
+    `mixed_precision` selects accelerate-native AMP for DDP/single-GPU runs and the FSDP2
+    `MixedPrecisionPolicy` for sharded runs (accelerate derives it). Sharded runs support
+    "no" and "bf16" only; fp16's GradScaler-over-DTensor path is unverified and fails fast
+    at config validation.
+    """
+
+    mixed_precision: str = "no"
+    gradient_accumulation: GradientAccumulationConfig = field(default_factory=GradientAccumulationConfig)
+    fsdp: FSDPConfig = field(default_factory=FSDPConfig)
+    ddp: DDPConfig = field(default_factory=DDPConfig)
+    compile: CompileConfig = field(default_factory=CompileConfig)
+    activation_checkpointing: ActivationCheckpointingConfig = field(
+        default_factory=ActivationCheckpointingConfig
+    )
+
+    def __post_init__(self) -> None:
+        """Validate the accelerate-facing scalar fields.
+
+        Raises:
+            ValueError: If ``mixed_precision`` is not one of ``"no"``, ``"fp16"``, ``"bf16"``.
+        """
+        if self.mixed_precision not in ("no", "fp16", "bf16"):
+            raise ValueError(
+                f"mixed_precision must be one of 'no', 'fp16', 'bf16', got {self.mixed_precision!r}."
+            )
+
+    def build(self, parallelism: ParallelismConfig, *, cpu: bool = False) -> "Accelerator":
+        """Translate the mirrored fields into a ready `Accelerator` (call once per process).
+
+        `parallelism` must already be resolved against the world size. The degradation matrix
+        is encoded here and nowhere else: sharded -> FSDP2 (+HSDP via the accelerate
+        `ParallelismConfig` mesh), replicated-only -> DDP kwargs, single process -> plain.
+
+        Args:
+            parallelism (ParallelismConfig): The resolved process topology; selects which
+                accelerate path (FSDP2 mesh, DDP kwargs handler, or plain) is configured.
+            cpu (bool): Force CPU execution even when CUDA is available. Defaults to False.
+
+        Returns:
+            Accelerator: The configured accelerate entry point for this process.
+        """
+        from accelerate import Accelerator
+
+        kwargs: dict = {
+            # LeRobot steps its scheduler manually once per training step; accelerate must not
+            # rescale scheduler stepping by num_processes.
+            "step_scheduler_with_optimizer": False,
+            "gradient_accumulation_plugin": self.gradient_accumulation.build_plugin(),
+            "mixed_precision": self.mixed_precision,
+            "cpu": cpu,
+        }
+        if parallelism.is_sharded:
+            kwargs["fsdp_plugin"] = self.fsdp.build_plugin()
+            kwargs["parallelism_config"] = _accelerate_parallelism_config(parallelism)
+        elif parallelism.is_replicated_only:
+            kwargs["kwargs_handlers"] = [self.ddp.build_kwargs_handler()]
+        return Accelerator(**kwargs)
+
+
+def _accelerate_parallelism_config(parallelism: ParallelismConfig) -> object:
+    """LeRobot topology -> accelerate `ParallelismConfig`.
+
+    CP is declared honestly (`cp_size = ring x ulysses`) so accelerate builds the canonical
+    mesh, folds CP into the FSDP shard group (`dp_shard_cp`), and duplicates batches within CP
+    groups. The ring/ulysses sub-structure stays private to `lerobot.distributed.ParallelDims`.
+
+    Args:
+        parallelism (ParallelismConfig): The resolved LeRobot topology to translate.
+
+    Returns:
+        object: The accelerate `ParallelismConfig` mirroring `dp_replicate`, `dp_shard`, and
+            the collapsed `cp_size` (annotated as `object` so importing this module never
+            imports accelerate).
+    """
+    from accelerate.parallelism_config import ParallelismConfig as AccelerateParallelismConfig
+
+    return AccelerateParallelismConfig(
+        dp_replicate_size=parallelism.dp_replicate,
+        dp_shard_size=parallelism.dp_shard,
+        cp_size=parallelism.cp_size,
+    )
@@ -14,6 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

+import logging
 from dataclasses import dataclass, field

 from lerobot.transforms import ImageTransformsConfig
@@ -21,6 +22,8 @@ from lerobot.utils.import_utils import get_safe_default_video_backend

 from .video import DEFAULT_DEPTH_UNIT, DEPTH_METER_UNIT, DEPTH_MILLIMETER_UNIT

+logger = logging.getLogger(__name__)
+

@dataclass
 class DatasetConfig:
@@ -29,10 +32,15 @@ class DatasetConfig:
    # "dataset_index" into the returned item. The index mapping is made according to the order in which the
    # datasets are provided.
    repo_id: str
+    # Hub repository type: "dataset" (default) or "bucket" for an HF Storage Bucket streamed over
+    # hf://buckets/. Buckets are streaming-only, so "bucket" requires streaming=true.
+    repo_type: str = "dataset"
    # Root directory for a concrete local dataset tree (e.g. 'dataset/path'). If None, local datasets are
    # looked up under $HF_LEROBOT_HOME/repo_id and Hub downloads use a revision-safe cache under $HF_LEROBOT_HOME/hub.
    root: str | None = None
    episodes: list[int] | None = None
+    # Episode indices to drop (e.g. corrupt or heterogeneous ones). Applied on top of `episodes`.
+    exclude_episodes: list[int] | None = None
    image_transforms: ImageTransformsConfig = field(default_factory=ImageTransformsConfig)
    revision: str | None = None
    use_imagenet_stats: bool = True
@@ -48,6 +56,16 @@ class DatasetConfig:
    eval_split: float = 0.0

    def __post_init__(self) -> None:
+        if self.repo_type not in ("dataset", "bucket"):
+            raise ValueError(f"repo_type must be 'dataset' or 'bucket', got {self.repo_type!r}")
+        if self.repo_type == "bucket" and not self.streaming:
+            raise ValueError(
+                "repo_type='bucket' is streaming-only: set streaming=true to train from an HF Storage Bucket."
+            )
+        if self.repo_type == "bucket" and self.eval_split != 0.0:
+            raise ValueError(
+                "eval_split requires map-style datasets and is not supported with repo_type='bucket'."
+            )
        if self.depth_output_unit not in (DEPTH_METER_UNIT, DEPTH_MILLIMETER_UNIT):
            raise ValueError(
                f"depth_output_unit must be '{DEPTH_METER_UNIT}' or '{DEPTH_MILLIMETER_UNIT}', got {self.depth_output_unit!r}"
@@ -62,6 +80,14 @@ class DatasetConfig:
            if len(self.episodes) != len(set(self.episodes)):
                duplicates = sorted({ep for ep in self.episodes if self.episodes.count(ep) > 1})
                raise ValueError(f"Episode indices contain duplicates: {duplicates}")
+        if self.exclude_episodes is not None:
+            negative_episodes = [episode for episode in self.exclude_episodes if episode < 0]
+            if negative_episodes:
+                logger.warning(
+                    "Ignoring negative exclude_episodes entries: %s",
+                    negative_episodes,
+                )
+                self.exclude_episodes = [episode for episode in self.exclude_episodes if episode >= 0]


@dataclass
@@ -93,9 +119,6 @@ class EvalConfig:
    recording_repo_id: str | None = None
    # Whether the pushed recording repositories should be private.
    recording_private: bool = False
-    # Whether to save the policy's imagined/predicted video (world-model policies only) as mp4s.
-    # Requests intermediate predictions from the policy each step; policies that produce none are unaffected.
-    save_predicted_video: bool = False

    def __post_init__(self) -> None:
        if self.recording_repo_id is not None and not self.recording:
@@ -116,6 +139,59 @@ class EvalConfig:
        return min(by_cpu, self.n_episodes, 64)


+@dataclass
+class EMAConfig:
+    """Exponential moving average (EMA) of the policy weights.
+
+    Standard practice for diffusion-style policies (Chi et al. 2023, "Diffusion Policy", section V.D):
+    the reference implementation enables it in every config and evaluates the EMA weights. Off by
+    default here because it keeps a second full copy of the parameters in memory.
+
+    The decay follows the warmup schedule from diffusers' `EMAModel`:
+    `decay_t = 1 - (1 + t / inv_gamma) ** -power`, clamped to `[min_decay, max_decay]`.
+    The defaults mirror the reference implementation. Alternatively, set `decay` for a constant
+    decay at every step, as used by openpi for pi0/pi05 (`ema_decay=0.99`).
+    """
+
+    enable: bool = False
+    # Constant decay coefficient (openpi-style, e.g. 0.99 for pi0/pi05). When set, the warmup
+    # schedule below is bypassed and the shadow uses this decay at every step.
+    decay: float | None = None
+    # Number of optimizer steps during which the shadow stays a hard copy of the live weights.
+    update_after_step: int = 0
+    # Warmup schedule parameters (see class docstring).
+    inv_gamma: float = 1.0
+    power: float = 0.75
+    min_decay: float = 0.0
+    max_decay: float = 0.9999
+    # Evaluate the EMA weights (instead of the live ones) during periodic env eval.
+    # Offline eval-loss (--eval_steps) always uses the live weights: it runs on every rank
+    # while the EMA shadow only lives on the main process.
+    use_for_eval: bool = True
+
+    def __post_init__(self) -> None:
+        if not (0.0 <= self.min_decay <= self.max_decay <= 1.0):
+            raise ValueError(
+                "Expected 0 <= ema.min_decay <= ema.max_decay <= 1, got "
+                f"min_decay={self.min_decay} and max_decay={self.max_decay}."
+            )
+        if self.inv_gamma <= 0:
+            raise ValueError(f"ema.inv_gamma must be positive, got {self.inv_gamma}.")
+        if self.power <= 0:
+            raise ValueError(f"ema.power must be positive, got {self.power}.")
+        if self.update_after_step < 0:
+            raise ValueError(f"ema.update_after_step must be >= 0, got {self.update_after_step}.")
+        if self.decay is not None:
+            if not 0.0 <= self.decay <= 1.0:
+                raise ValueError(f"ema.decay must be in [0, 1], got {self.decay}.")
+            # Keep the literals in sync with the field defaults above.
+            if self.min_decay != 0.0 or self.max_decay != 0.9999:
+                raise ValueError(
+                    "ema.decay (constant decay) and ema.min_decay/ema.max_decay (schedule clamp) are "
+                    "mutually exclusive: set one or the other."
+                )
+
+
@dataclass
 class PeftConfig:
    # PEFT offers many fine-tuning methods, layer adapters being the most common and currently also the most
@@ -0,0 +1,190 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Declarative process topology for distributed training and inference.
+
+The mesh convention (canonical row-major rank layout, outermost first)::
+
+    (dp_replicate, dp_shard, ring, ulysses)
+
+- ``dp_replicate x dp_shard`` is the data-parallel world: HSDP replicates over
+  ``dp_replicate`` and shards parameters over ``dp_shard``. FSDP2's actual shard
+  group folds context parallelism in (``dp_shard x ring x ulysses``), matching
+  accelerate's ``dp_shard_cp`` flattening and torchtitan's ``fsdp`` axis.
+- ``ring`` is the outer and ``ulysses`` the inner context-parallel dim
+  (diffusers convention: ulysses all-to-all exchanges run over adjacent, typically
+  NVLink-connected ranks).
+- ``cfg_parallel`` (classifier-free-guidance parallelism) is a branch-parallel,
+  inference-only dim that sits between dp and the sequence dims. It never
+  affects weight sharding or checkpoints.
+
+This module is pure configuration: plain-typed dataclasses that draccus can
+round-trip through the CLI and ``train_config.json``. Runtime objects (device
+meshes, process groups) live in :mod:`lerobot.distributed`.
+"""
+
+import os
+from dataclasses import dataclass, field
+
+
+@dataclass
+class ContextParallelConfig:
+    """Ring x Ulysses context parallelism (sequence parallelism for attention).
+
+    Both degrees are configured placeholders in this release: the CP engine is not implemented
+    yet, and enabling either degree > 1 fails fast at config validation. The fields exist now so
+    that the CLI surface, checkpoint metadata, and mesh math are stable when the engine lands.
+    """
+
+    ring_degree: int = 1
+    ulysses_degree: int = 1
+
+    def __post_init__(self) -> None:
+        """Validate the declared context-parallel degrees.
+
+        Raises:
+            ValueError: If ``ring_degree`` or ``ulysses_degree`` is < 1.
+        """
+        if self.ring_degree < 1 or self.ulysses_degree < 1:
+            raise ValueError(
+                f"Context-parallel degrees must be >= 1, got ring_degree={self.ring_degree}, "
+                f"ulysses_degree={self.ulysses_degree}."
+            )
+
+    @property
+    def size(self) -> int:
+        """Total number of ranks a full sequence is sharded across."""
+        return self.ring_degree * self.ulysses_degree
+
+
+@dataclass
+class ParallelismConfig:
+    """Degrees of every parallelism dim. Invariant: their product equals the world size.
+
+    Degradations are expressed purely through the degrees (no mode flags):
+
+    - single process: all degrees 1;
+    - DDP: ``dp_replicate == world_size`` (auto-filled when every sharding field is left at its
+      default — plain ``torchrun`` keeps today's out-of-the-box behavior);
+    - FSDP: ``dp_shard > 1`` (or ``-1`` to fill the remaining world into the shard dim);
+    - HSDP: ``dp_replicate > 1`` and ``dp_shard > 1``.
+
+    ``resolve()`` turns the declared degrees into concrete ones once the world size is known and
+    is the single place the world-size equation is enforced. It is called by
+    :func:`lerobot.distributed.factory.make_accelerator`; the config is inert until then.
+    """
+
+    dp_replicate: int = 1
+    # -1 is an explicit opt-in sentinel: shard over world_size // (dp_replicate * cp).
+    dp_shard: int = 1
+    context_parallel: ContextParallelConfig = field(default_factory=ContextParallelConfig)
+    # Classifier-free-guidance parallelism — inference-only (cosmos/vllm-omni precedent:
+    # cond/uncond branches on different ranks). Reserved for the serving round; training
+    # validates it to 1. Meaningful values are 1 or 2 (Cosmos3 has two CFG branches).
+    cfg_parallel: int = 1
+
+    def __post_init__(self) -> None:
+        """Validate the declared degrees (world-size-independent checks only).
+
+        Raises:
+            ValueError: If ``dp_replicate`` is < 1, ``dp_shard`` is neither >= 1 nor the
+                ``-1`` infer sentinel, or ``cfg_parallel`` is not 1 or 2.
+        """
+        if self.dp_replicate < 1:
+            raise ValueError(f"dp_replicate must be >= 1, got {self.dp_replicate}.")
+        if self.dp_shard < 1 and self.dp_shard != -1:
+            raise ValueError(f"dp_shard must be >= 1, or -1 to infer, got {self.dp_shard}.")
+        if self.cfg_parallel not in (1, 2):
+            raise ValueError(f"cfg_parallel must be 1 or 2, got {self.cfg_parallel}.")
+
+    @property
+    def cp_size(self) -> int:
+        """Total context-parallel size (``ring_degree * ulysses_degree``)."""
+        return self.context_parallel.size
+
+    @property
+    def is_sharded(self) -> bool:
+        """True when the run uses FSDP2 (parameters sharded); selects the sharded engine path."""
+        return self.dp_shard != 1 or self.cp_size > 1
+
+    @property
+    def is_replicated_only(self) -> bool:
+        """True for plain DDP (weights replicated, no sharding)."""
+        return not self.is_sharded and self.dp_replicate > 1
+
+    @property
+    def dp_world_size(self) -> int:
+        """Number of distinct data-parallel workers (batches are sharded this many ways).
+
+        Returns:
+            int: ``dp_replicate * dp_shard``.
+
+        Raises:
+            RuntimeError: If accessed while ``dp_shard`` is still the ``-1`` sentinel, i.e.
+                before :meth:`resolve` has bound the degrees to a world size.
+        """
+        if self.dp_shard == -1:
+            raise RuntimeError("dp_world_size is undefined before resolve() fills dp_shard=-1.")
+        return self.dp_replicate * self.dp_shard
+
+    def resolve(self, world_size: int) -> None:
+        """Bind the declared degrees to a concrete world size (idempotent).
+
+        Fills the ``dp_shard=-1`` sentinel, auto-fills ``dp_replicate`` for the DDP degradation,
+        and enforces ``dp_replicate * dp_shard * cp == world_size`` with every degree echoed on
+        failure.
+
+        Args:
+            world_size (int): Total number of launched processes (torchrun's ``WORLD_SIZE``).
+
+        Raises:
+            ValueError: If a context-parallel degree is > 1 (the CP engine is not implemented
+                yet), if ``dp_shard=-1`` cannot be inferred because ``world_size`` is not
+                divisible by ``dp_replicate * cp``, or if the resolved degrees do not multiply
+                to ``world_size``.
+        """
+        if self.cp_size > 1:
+            raise ValueError(
+                "Context parallelism is not implemented yet: ring_degree and ulysses_degree "
+                "must be 1. The fields are reserved for the CP engine round."
+            )
+        if self.is_sharded:
+            if self.dp_shard == -1:
+                self.dp_shard, remainder = divmod(world_size, self.dp_replicate * self.cp_size)
+                if remainder or self.dp_shard < 1:
+                    raise ValueError(
+                        f"Cannot infer dp_shard: world_size={world_size} is not divisible by "
+                        f"dp_replicate={self.dp_replicate} * cp={self.cp_size}."
+                    )
+        elif self.dp_replicate == 1:
+            # Untouched config on a multi-process launch: fill the DDP degradation.
+            self.dp_replicate = world_size
+        total = self.dp_replicate * self.dp_shard * self.cp_size
+        if total != world_size:
+            raise ValueError(
+                f"Parallelism degrees do not multiply to the world size: dp_replicate="
+                f"{self.dp_replicate} * dp_shard={self.dp_shard} * ring="
+                f"{self.context_parallel.ring_degree} * ulysses="
+                f"{self.context_parallel.ulysses_degree} = {total} != WORLD_SIZE={world_size}."
+            )
+
+
+def world_size_from_env() -> int:
+    """World size as set by torchrun (or 1 outside distributed launches).
+
+    Returns:
+        int: The ``WORLD_SIZE`` environment variable, or 1 when unset.
+    """
+    return int(os.environ.get("WORLD_SIZE", "1"))
@@ -23,6 +23,7 @@ from typing import Any, Literal, get_args

 MessageRole = Literal["user", "assistant", "system", "tool"]
 MessageStream = Literal["high_level", "low_level"]
+RecipeRoute = Literal["vqa"]

 DEFAULT_BINDINGS = {
    "subtask": "active_at(t, style=subtask)",
@@ -40,6 +41,7 @@ discovery (here) and rendered-message substitution (in ``language_render``)."""

 _VALID_ROLES = frozenset(get_args(MessageRole))
 _VALID_STREAMS = frozenset(get_args(MessageStream))
+_VALID_ROUTES = frozenset(get_args(RecipeRoute))


@dataclass
@@ -78,7 +80,7 @@ class MessageTurn:
            raise ValueError(f"Unsupported message stream: {self.stream!r}")
        if self.content is None and self.tool_calls_from is None:
            raise ValueError("MessageTurn.content is required unless tool_calls_from is set.")
-        if self.content is not None and not isinstance(self.content, (str, list)):
+        if self.content is not None and not isinstance(self.content, str | list):
            raise TypeError("MessageTurn.content must be a string, a list of HF-style blocks, or None.")
        if isinstance(self.content, list):
            for block in self.content:
@@ -99,13 +101,16 @@ class TrainingRecipe:

    A recipe is either a *message recipe* (``messages`` plus optional
    ``bindings``) or a *blend recipe* (``blend`` mapping names to weighted
-    sub-recipes). ``weight`` is only meaningful inside a blend.
+    sub-recipes). ``weight`` and ``route`` are only meaningful inside a blend;
+    ``route: vqa`` gives sparse VQA annotations priority over normal weighted
+    selection.
    """

    messages: list[MessageTurn] | None = None
    bindings: dict[str, str] | None = None
    blend: dict[str, TrainingRecipe] | None = None
    weight: float | None = None
+    route: RecipeRoute | None = None

    def __post_init__(self) -> None:
        """Validate that exactly one of ``messages`` or ``blend`` is set."""
@@ -113,6 +118,10 @@ class TrainingRecipe:
            raise ValueError("TrainingRecipe must set only one of messages or blend.")
        if self.messages is None and self.blend is None:
            raise ValueError("TrainingRecipe must set one of messages or blend.")
+        if self.route is not None and self.route not in _VALID_ROUTES:
+            raise ValueError(f"Unsupported recipe route: {self.route!r}")
+        if self.blend is not None and self.route is not None:
+            raise ValueError("TrainingRecipe.route may only be set on a message recipe inside a blend.")

        if self.messages is not None:
            self._validate_message_recipe()
@@ -147,8 +156,9 @@ class TrainingRecipe:
        return cls.from_dict(data)

    def _validate_message_recipe(self) -> None:
-        """Ensure every templated binding is known and at least one turn is a target."""
-        assert self.messages is not None
+        """Validate bindings and require text or low-level action supervision."""
+        if self.messages is None:
+            raise ValueError("Cannot validate a message recipe without messages.")
        known_bindings = set(DEFAULT_BINDINGS) | set(self.bindings or {}) | {"task"}

        for turn in self.messages:
@@ -156,12 +166,19 @@ class TrainingRecipe:
            if missing:
                raise ValueError(f"MessageTurn references unknown binding(s): {sorted(missing)}")

-        if not any(turn.target for turn in self.messages):
-            raise ValueError("Message recipes must contain at least one target turn.")
+        has_target = any(turn.target for turn in self.messages)
+        has_low_level = any(turn.stream == "low_level" for turn in self.messages)
+        if not (has_target or has_low_level):
+            raise ValueError(
+                "Message recipes must contain at least one supervised turn — "
+                "either ``target: true`` (text CE) or ``stream: low_level`` "
+                "(flow/action loss)."
+            )

    def _validate_blend_recipe(self) -> None:
        """Ensure each blend component is a non-empty, weighted message recipe."""
-        assert self.blend is not None
+        if self.blend is None:
+            raise ValueError("Cannot validate a blend recipe without blend components.")
        if not self.blend:
            raise ValueError("Blend recipes must contain at least one component.")

@@ -0,0 +1,16 @@
+# Predicts subtasks from tasks and trains subtask-conditioned action flow without memory or plans.
+# Requires `subtask` annotations; samples with missing `if_present` bindings do not render.
+
+blend:
+
+  high_level_subtask:
+    weight: 0.30
+    messages:
+      - {role: user, content: "${task}", stream: high_level}
+      - {role: assistant, content: "${subtask}", stream: high_level, target: true, if_present: subtask}
+
+  low_level_execution:
+    weight: 0.70
+    messages:
+      # The low-level stream trains action flow on the generated or annotated subtask.
+      - {role: user, content: "${subtask}", stream: low_level, if_present: subtask}
@@ -0,0 +1,13 @@
+# Paper-style joint sequence (pi0.5 §IV-B): one sample supervises the subtask
+# text with CE and, because the assistant turn is part of the prefix, conditions
+# the FAST and flow action losses on the same annotated subtask in one forward.
+# The supervised span is attended causally; the action losses see task + subtask.
+#
+# Pair with `--policy.joint_subtask_conditioning=true` at inference so the flow
+# prefix reproduces this layout (task turn with state + causal generated subtask).
+# Samples without a `subtask` annotation fall back to a plain task-prompt
+# low-level sample via `if_present`.
+
+messages:
+  - {role: user, content: "${task}", stream: low_level}
+  - {role: assistant, content: "${subtask}", stream: low_level, target: true, if_present: subtask}
@@ -0,0 +1,30 @@
+# Trains subtask prediction, subtask-conditioned action flow, and memory updates without plans.
+# Requires `subtask` and `memory`; missing `if_present` bindings skip the affected sub-recipe.
+
+blend:
+
+  high_level_subtask:
+    weight: 0.25
+    messages:
+      - {role: user, content: "${task}", stream: high_level}
+      - {role: assistant, content: "${subtask}", stream: high_level, target: true, if_present: subtask}
+
+  low_level_execution:
+    weight: 0.60
+    messages:
+      # The low-level stream trains action flow on the generated or annotated subtask.
+      - {role: user, content: "${subtask}", stream: low_level, if_present: subtask}
+
+  memory_update:
+    # `active_at` densifies sparse boundaries while preserving the prior-memory/subtask mapping.
+    # Inference controls update timing through `subtask_change` events.
+    weight: 0.15
+    bindings:
+      prior_memory: "nth_prev(style=memory, offset=1)"
+      current_memory: "active_at(t, style=memory)"
+      completed_subtask: "nth_prev(style=subtask, offset=1)"
+    messages:
+      - {role: user, content: "${task}", stream: high_level}
+      - {role: assistant, content: "Previous memory: ${prior_memory}", stream: high_level, if_present: prior_memory}
+      - {role: user, content: "Completed subtask: ${completed_subtask}", stream: high_level, if_present: completed_subtask}
+      - {role: assistant, content: "${current_memory}", stream: high_level, target: true, if_present: current_memory}
@@ -0,0 +1,72 @@
+# Adds memory, spoken interjection responses, and camera-grounded VQA to subtask/action training.
+# Missing optional annotations skip only their sub-recipe; `say` tool calls tokenize as `<say>...</say>`.
+
+blend:
+
+  high_level_subtask:
+    weight: 0.25
+    messages:
+      - {role: user, content: "${task}", stream: high_level}
+      - {role: assistant, content: "${subtask}", stream: high_level, target: true, if_present: subtask}
+
+  low_level_execution:
+    weight: 0.40
+    messages:
+      # The low-level stream trains action flow on the generated or annotated subtask.
+      - {role: user, content: "${subtask}", stream: low_level, if_present: subtask}
+
+  memory_update:
+    # `active_at` densifies sparse boundaries while preserving the prior-memory/subtask mapping.
+    # Inference controls update timing through `subtask_change` events.
+    weight: 0.10
+    bindings:
+      prior_memory: "nth_prev(style=memory, offset=1)"
+      current_memory: "active_at(t, style=memory)"
+      completed_subtask: "nth_prev(style=subtask, offset=1)"
+    messages:
+      - {role: user, content: "${task}", stream: high_level}
+      - {role: assistant, content: "Previous memory: ${prior_memory}", stream: high_level, if_present: prior_memory}
+      - {role: user, content: "Completed subtask: ${completed_subtask}", stream: high_level, if_present: completed_subtask}
+      - {role: assistant, content: "${current_memory}", stream: high_level, target: true, if_present: current_memory}
+
+  user_interjection_response:
+    weight: 0.10
+    bindings:
+      interjection: "emitted_at(t, style=interjection)"
+      speech: "emitted_at(t, role=assistant, tool_name=say)"
+    messages:
+      - {role: user, content: "${task}", stream: high_level}
+      - {role: user, content: "${interjection}", stream: high_level, if_present: interjection}
+      # The assistant target is a `say` tool call flattened to a `<say>...</say>` marker.
+      - {role: assistant, stream: high_level, target: true, if_present: speech, tool_calls_from: speech}
+
+  # Each camera uses a separate VQA sub-recipe for view-specific binding.
+  ask_vqa_top:
+    weight: 0.075
+    route: vqa
+    bindings:
+      vqa_query: "emitted_at(t, style=vqa, role=user, camera=observation.images.front)"
+      vqa: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.front)"
+    messages:
+      - role: user
+        stream: high_level
+        if_present: vqa_query
+        content:
+          - {type: image, feature: observation.images.front}
+          - {type: text, text: "${vqa_query}"}
+      - {role: assistant, content: "${vqa}", stream: high_level, target: true, if_present: vqa}
+
+  ask_vqa_wrist:
+    weight: 0.075
+    route: vqa
+    bindings:
+      vqa_query: "emitted_at(t, style=vqa, role=user, camera=observation.images.wrist)"
+      vqa: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.wrist)"
+    messages:
+      - role: user
+        stream: high_level
+        if_present: vqa_query
+        content:
+          - {type: image, feature: observation.images.wrist}
+          - {type: text, text: "${vqa_query}"}
+      - {role: assistant, content: "${vqa}", stream: high_level, target: true, if_present: vqa}
@@ -18,6 +18,7 @@ import multiprocessing
 import os
 import tempfile
 from dataclasses import dataclass, field
+from enum import Enum
 from pathlib import Path
 from typing import Any

@@ -26,19 +27,49 @@ from huggingface_hub import hf_hub_download
 from huggingface_hub.errors import HfHubHTTPError

 from lerobot import envs
+from lerobot.configs.accelerator import AcceleratorConfig, ActivationCheckpointingMode
+from lerobot.configs.parallelism import ParallelismConfig
 from lerobot.optim import LRSchedulerConfig, OptimizerConfig
 from lerobot.utils.constants import PRETRAINED_MODEL_DIR
 from lerobot.utils.hub import HubMixin, find_latest_hub_checkpoint
 from lerobot.utils.sample_weighting import SampleWeightingConfig

 from . import parser
-from .default import DatasetConfig, EvalConfig, JobConfig, PeftConfig, WandBConfig
+from .default import DatasetConfig, EMAConfig, EvalConfig, JobConfig, PeftConfig, WandBConfig
 from .policies import PreTrainedConfig
 from .rewards import RewardModelConfig

 TRAIN_CONFIG_NAME = "train_config.json"


+class CheckpointFormat(str, Enum):
+    """Model-artifact format inside training checkpoints.
+
+    Selects only the *model* artifact; the training_state layout is format-independent (the
+    optimizer channel is always DCP under sharded runs, safetensors+json otherwise).
+
+    - SAFETENSORS (default): a full `model.safetensors` — maximum compatibility, one gather per
+      save under sharding.
+    - DCP: sharded `pytorch_model_fsdp_0/*.distcp` only — fastest save/resume; convert with
+      `lerobot-convert-dcp` before distributing.
+    - SAFETENSORS_AND_DCP: both artifacts, written independently.
+    """
+
+    SAFETENSORS = "safetensors"
+    DCP = "dcp"
+    SAFETENSORS_AND_DCP = "safetensors_dcp"
+
+    @property
+    def wants_safetensors(self) -> bool:
+        """True when a full `model.safetensors` artifact should be written."""
+        return self in (CheckpointFormat.SAFETENSORS, CheckpointFormat.SAFETENSORS_AND_DCP)
+
+    @property
+    def wants_dcp(self) -> bool:
+        """True when sharded DCP model shards (`pytorch_model_fsdp_0/`) should be written."""
+        return self in (CheckpointFormat.DCP, CheckpointFormat.SAFETENSORS_AND_DCP)
+
+
 def _migrate_legacy_rabc_fields(config: dict[str, Any]) -> dict[str, Any] | None:
    """Return migrated payload for legacy RA-BC fields, or None when no migration is needed."""
    legacy_fields = (
@@ -121,10 +152,19 @@ class TrainPipelineConfig(HubMixin):
    # Checkpoint is saved every `save_freq` training iterations and after the last training step.
    # A non-positive value disables periodic saving, keeping only the final checkpoint.
    save_freq: int = 20_000
+    # Model-artifact format inside checkpoints; non-default values require a sharded run.
+    checkpoint_format: CheckpointFormat = CheckpointFormat.SAFETENSORS
    use_policy_training_preset: bool = True
    optimizer: OptimizerConfig | None = None
    scheduler: LRSchedulerConfig | None = None
+    # Process topology: dp_replicate / dp_shard (HSDP) and context-parallel degree placeholders.
+    parallelism: ParallelismConfig = field(default_factory=ParallelismConfig)
+    # Execution runtime handed to the Accelerator: mixed precision, gradient accumulation,
+    # FSDP/DDP tuning knobs, compile & activation-checkpointing placeholders.
+    accelerator: AcceleratorConfig = field(default_factory=AcceleratorConfig)
    eval: EvalConfig = field(default_factory=EvalConfig)
+    # Maintain an EMA shadow of the policy weights during training (see EMAConfig).
+    ema: EMAConfig = field(default_factory=EMAConfig)
    wandb: WandBConfig = field(default_factory=WandBConfig)
    peft: PeftConfig | None = None

@@ -291,6 +331,60 @@ class TrainPipelineConfig(HubMixin):
        if self.save_checkpoint_to_hub and not (self.policy is not None and self.policy.repo_id):
            raise ValueError("save_checkpoint_to_hub requires --policy.repo_id.")

+        self._validate_distributed()
+
+    def _validate_distributed(self) -> None:
+        """Fail-fasts for the distributed-training scope.
+
+        Raises:
+            ValueError: If the config requests anything outside the verified scope: context
+                parallelism or CFG parallelism (reserved placeholders), the compile or
+                activation-checkpointing placeholders, a DCP checkpoint format on a
+                non-sharded run, or — under sharded training — fp16 mixed precision, PEFT,
+                reward-model training, in-training environment evaluation, or multi-optimizer
+                configs.
+        """
+        if self.parallelism.cp_size > 1:
+            raise ValueError(
+                "Context parallelism is not implemented yet: --parallelism.context_parallel.* "
+                "degrees must be 1 (reserved for the CP engine round)."
+            )
+        if self.parallelism.cfg_parallel != 1:
+            raise ValueError(
+                "CFG parallelism is inference-only and must be 1 for training "
+                "(cfg_parallel is reserved for the serving round)."
+            )
+        if self.accelerator.compile.enabled:
+            raise ValueError("--accelerator.compile is a placeholder and not wired yet.")
+        if self.accelerator.activation_checkpointing.mode is not ActivationCheckpointingMode.NONE:
+            raise ValueError("--accelerator.activation_checkpointing is a placeholder and not wired yet.")
+        if self.checkpoint_format is not CheckpointFormat.SAFETENSORS and not self.parallelism.is_sharded:
+            raise ValueError(
+                f"checkpoint_format={self.checkpoint_format.value} requires a sharded run "
+                "(--parallelism.dp_shard != 1); non-sharded checkpoints are always safetensors."
+            )
+        if self.parallelism.is_sharded:
+            if self.accelerator.mixed_precision == "fp16":
+                raise ValueError(
+                    "fp16 is not supported under sharded training (GradScaler over DTensor "
+                    "gradients is unverified); use bf16 or full precision."
+                )
+            if self.peft is not None:
+                raise ValueError("PEFT is not supported under sharded training yet.")
+            if self.is_reward_model_training:
+                raise ValueError(
+                    "Reward-model training is not supported under sharded training yet "
+                    "(reward models declare no FSDP wrap units and have no sharded save path)."
+                )
+            if self.env is not None and self.env_eval_freq > 0:
+                raise ValueError(
+                    "In-training environment evaluation is not supported under sharded training "
+                    "(a rank-0-only rollout of a sharded model deadlocks on collectives); set "
+                    "--env_eval_freq=0 and evaluate with lerobot-eval on saved checkpoints."
+                )
+            if self.optimizer is not None and self.optimizer.builds_multiple_optimizers:
+                raise ValueError("Multi-optimizer configs are not supported under sharded training.")
+
    @classmethod
    def __get_path_fields__(cls) -> list[str]:
        """Keys for draccus pretrained-path loading."""
@@ -76,7 +76,7 @@ import torch
 from pydantic import BaseModel, Field
 from transformers import AutoProcessor, Qwen3VLMoeForConditionalGeneration

-from lerobot.datasets import LeRobotDataset
+from lerobot.datasets import LeRobotDataset, resolve_episode_indices


 # Pydantic Models for SARM Subtask Annotation
@@ -1049,7 +1049,10 @@ def main():
    torch_dtype = {"bfloat16": torch.bfloat16, "float16": torch.float16, "float32": torch.float32}[args.dtype]

    # Determine episodes
-    episode_indices = args.episodes or list(range(dataset.meta.total_episodes))
+    resolved_episodes = resolve_episode_indices(args.episodes, dataset.meta.total_episodes)
+    episode_indices = (
+        resolved_episodes if resolved_episodes is not None else list(range(dataset.meta.total_episodes))
+    )

    existing_annotations = load_annotations_from_dataset(dataset.root, prefix="sparse")
    if args.skip_existing:
@@ -52,7 +52,7 @@ from .pipeline_features import aggregate_pipeline_dataset_features, create_initi
 from .pyav_utils import check_video_encoder_parameters_pyav, detect_available_encoders_pyav
 from .sampler import EpisodeAwareSampler, compute_sampler_state
 from .streaming_dataset import StreamingLeRobotDataset
-from .utils import DEFAULT_EPISODES_PATH, create_lerobot_dataset_card
+from .utils import DEFAULT_EPISODES_PATH, create_lerobot_dataset_card, resolve_episode_indices
 from .video_utils import VideoEncodingManager

 # NOTE: Low-level I/O functions (cast_stats_to_numpy, get_parquet_file_size_in_mb, etc.)
@@ -97,6 +97,7 @@ __all__ = [
    "reencode_dataset",
    "remove_feature",
    "resolve_delta_timestamps",
+    "resolve_episode_indices",
    "safe_stop_image_writer",
    "split_dataset",
    "write_stats",
@@ -22,6 +22,7 @@ from pathlib import Path
 from typing import Any, NotRequired, TypedDict

 import datasets
+import numpy as np
 import pandas as pd
 import tqdm

@@ -303,6 +304,46 @@ def update_meta_data(
    df["dataset_to_index"] = df["dataset_to_index"] + dst_meta.info.total_frames
    df["episode_index"] = df["episode_index"] + dst_meta.info.total_episodes

+    # Per-episode stats still describe the pre-merge values of the bookkeeping columns
+    # reindexed above. index/episode_index shift by a constant; task_index is relabeled,
+    # so recompute it from the episode's (stable) task strings via the unified tasks table.
+    shift_stat_keys = ("min", "max", "mean", "q01", "q10", "q50", "q90", "q99")
+    for name, offset in (
+        ("episode_index", dst_meta.info.total_episodes),
+        ("index", dst_meta.info.total_frames),
+    ):
+        for stat in shift_stat_keys:
+            col = f"stats/{name}/{stat}"
+            if col in df.columns:
+                df[col] = df[col] + offset
+
+    if any(c.startswith("stats/task_index/") for c in df.columns):
+        quantiles = {"q01": 0.01, "q10": 0.10, "q50": 0.50, "q90": 0.90, "q99": 0.99}
+        ids_per_row = [
+            np.array([dst_meta.tasks.loc[t, "task_index"] for t in tasks], dtype=np.float64)
+            for tasks in df["tasks"]
+        ]
+
+        def _task_stat(ids, stat):
+            if stat == "min":
+                return ids.min()
+            if stat == "max":
+                return ids.max()
+            if stat == "std":
+                return ids.std()
+            if stat in quantiles:
+                return np.quantile(ids, quantiles[stat])
+            return ids.mean()
+
+        for stat in ("min", "max", "mean", "std", *quantiles):
+            col = f"stats/task_index/{stat}"
+            if col in df.columns:
+                # np.full_like preserves each cell container and dtype so the parquet schema is unchanged.
+                df[col] = [
+                    np.full_like(orig, _task_stat(ids, stat))
+                    for orig, ids in zip(df[col], ids_per_row, strict=True)
+                ]
+
    return df


@@ -613,8 +613,15 @@ def aggregate_feature_stats(stats_ft_list: list[dict[str, dict]]) -> dict[str, d
        for q_key in quantile_keys:
            if all(q_key in s for s in stats_ft_list):
                quantile_values = np.stack([s[q_key] for s in stats_ft_list])
-                weighted_quantiles = quantile_values * counts
-                aggregated[q_key] = weighted_quantiles.sum(axis=0) / total_count
+                # Exact global quantiles cannot be recovered from quantile summaries.
+                # Keep a conservative envelope of the available estimates: min
+                # for lower quantiles and max for upper quantiles. The resulting
+                # values are bounds across the inputs, not global quantile estimates.
+                q_percent = int(q_key[1:])
+                if q_percent <= 50:
+                    aggregated[q_key] = np.min(quantile_values, axis=0)
+                else:
+                    aggregated[q_key] = np.max(quantile_values, axis=0)

    return aggregated

@@ -39,6 +39,7 @@ from .io_utils import (
    hf_transform_to_torch,
    load_nested_dataset,
 )
+from .utils import resolve_episode_indices
 from .video_utils import decode_video_frames


@@ -83,7 +84,7 @@ class DatasetReader:
        """
        self._meta = meta
        self.root = root
-        self.episodes = episodes
+        self.episodes = resolve_episode_indices(episodes, meta.total_episodes)
        self._tolerance_s = tolerance_s
        self._video_backend = video_backend
        if image_transforms is not None and not callable(image_transforms):
@@ -163,10 +164,34 @@ class DatasetReader:
    def _load_hf_dataset(self) -> datasets.Dataset:
        """hf_dataset contains all the observations, states, actions, rewards, etc."""
        features = get_hf_features_from_features(self._meta.features)
+        self._validate_language_columns_declared(features)
        hf_dataset = load_nested_dataset(self.root / "data", features=features, episodes=self.episodes)
        hf_dataset.set_transform(hf_transform_to_torch)
        return hf_dataset

+    def _validate_language_columns_declared(self, features: datasets.Features) -> None:
+        """Require language columns stored in Parquet to be declared in metadata."""
+        # Leave empty datasets to fail through the normal loading path.
+        try:
+            sample = next((self.root / "data").glob("*/*.parquet"))
+        except StopIteration:
+            return
+
+        from pyarrow import parquet as _pq  # noqa: PLC0415
+
+        # LeRobot shards are schema-uniform, so one schema represents the dataset.
+        schema_names = set(_pq.read_schema(sample).names)
+        from .language import LANGUAGE_COLUMNS  # noqa: PLC0415
+
+        missing = sorted(set(LANGUAGE_COLUMNS) & schema_names - set(features))
+        if missing:
+            raise ValueError(
+                f"Dataset Parquet files contain language feature(s) missing from metadata: {missing}. "
+                "Metadata must describe the stored data; add the entries returned by "
+                "lerobot.datasets.language.language_feature_info() to meta/info.json['features'] "
+                "or rerun the annotation pipeline's metadata synchronization."
+            )
+
    def _check_cached_episodes_sufficient(self) -> bool:
        """Check if the cached dataset contains all requested episodes and their video files."""
        if self.hf_dataset is None or len(self.hf_dataset) == 0:
@@ -29,6 +29,7 @@ from .dataset_metadata import LeRobotDatasetMetadata
 from .lerobot_dataset import LeRobotDataset
 from .multi_dataset import MultiLeRobotDataset
 from .streaming_dataset import StreamingLeRobotDataset
+from .utils import resolve_episode_indices


 def resolve_delta_timestamps(
@@ -84,14 +85,24 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas

    if isinstance(cfg.dataset.repo_id, str):
        ds_meta = LeRobotDatasetMetadata(
-            cfg.dataset.repo_id, root=cfg.dataset.root, revision=cfg.dataset.revision
+            cfg.dataset.repo_id,
+            root=cfg.dataset.root,
+            revision=cfg.dataset.revision,
+            repo_type=cfg.dataset.repo_type,
        )
        delta_timestamps = resolve_delta_timestamps(cfg.trainable_config, ds_meta)
+        episodes = resolve_episode_indices(
+            cfg.dataset.episodes, ds_meta.total_episodes, cfg.dataset.exclude_episodes
+        )
        if not cfg.dataset.streaming:
+            if cfg.dataset.repo_type == "bucket":
+                raise ValueError(
+                    "repo_type='bucket' is streaming-only: set dataset.streaming=true to train from an HF Storage Bucket."
+                )
            dataset = LeRobotDataset(
                cfg.dataset.repo_id,
                root=cfg.dataset.root,
-                episodes=cfg.dataset.episodes,
+                episodes=episodes,
                delta_timestamps=delta_timestamps,
                image_transforms=image_transforms,
                revision=cfg.dataset.revision,
@@ -104,13 +115,14 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas
            dataset = StreamingLeRobotDataset(
                cfg.dataset.repo_id,
                root=cfg.dataset.root,
-                episodes=cfg.dataset.episodes,
+                episodes=episodes,
                delta_timestamps=delta_timestamps,
                image_transforms=image_transforms,
                revision=cfg.dataset.revision,
                max_num_shards=cfg.num_workers,
                tolerance_s=cfg.tolerance_s,
                return_uint8=True,
+                repo_type=cfg.dataset.repo_type,
            )
    else:
        raise NotImplementedError("The MultiLeRobotDataset isn't supported for now.")
@@ -162,14 +162,32 @@ def render_sample(
    task: str | None = None,
    dataset_ctx: Any | None = None,
 ) -> RenderedMessages | None:
-    """Render the chat-style messages for a single dataset sample.
+    """Render recipe-defined messages and supervision for one dataset sample.

-    Resolves the recipe's bindings against ``persistent`` and ``events`` rows
-    at frame timestamp ``t``, then expands the recipe's message templates.
-    Returns ``None`` if the resolved sample contains no target message.
+    Resolves bindings against ``persistent`` and ``events`` at frame timestamp
+    ``t``. Blend recipes first route matching sparse VQA annotations, then use
+    deterministic weighted selection for the remaining samples. Returns
+    ``None`` when the selected recipe provides no text or low-level action
+    supervision for this sample.
    """
    persistent_rows = _normalize_rows(persistent or [])
    event_rows = _normalize_rows(events or [])
+
+    # Route sparse VQA frames to a matching view-specific component before weighted selection.
+    # This avoids dropping annotated frames or selecting VQA without annotations.
+    if recipe.blend is not None:
+        vqa_rendered = _render_vqa_if_present(
+            recipe,
+            persistent=persistent_rows,
+            events=event_rows,
+            t=t,
+            sample_idx=sample_idx,
+            task=task,
+            dataset_ctx=dataset_ctx,
+        )
+        if vqa_rendered is not None:
+            return vqa_rendered
+
    selected_recipe = _select_recipe(recipe, sample_idx)
    bindings = _resolve_bindings(
        selected_recipe,
@@ -183,6 +201,58 @@ def render_sample(
    return _render_message_recipe(selected_recipe, bindings)


+def _render_vqa_if_present(
+    recipe: TrainingRecipe,
+    *,
+    persistent: Sequence[LanguageRow],
+    events: Sequence[LanguageRow],
+    t: float,
+    sample_idx: int,
+    task: str | None,
+    dataset_ctx: Any | None,
+) -> RenderedMessages | None:
+    """Render a matching VQA component, or return ``None`` for normal selection.
+
+    Multiple matching views are selected deterministically by relative weight.
+    """
+    if recipe.blend is None:
+        return None
+    renderable: list[tuple[float, RenderedMessages]] = []
+    for component in recipe.blend.values():
+        if component.route != "vqa":
+            continue
+        bindings = _resolve_bindings(
+            component,
+            persistent=persistent,
+            events=events,
+            t=t,
+            sample_idx=sample_idx,
+            task=task,
+            dataset_ctx=dataset_ctx,
+        )
+        rendered = _render_message_recipe(component, bindings)
+        if rendered is not None:
+            if component.weight is None:
+                raise ValueError("Routed VQA blend components must define a weight.")
+            renderable.append((component.weight, rendered))
+
+    if not renderable:
+        return None
+    if len(renderable) == 1:
+        return renderable[0][1]
+
+    # Choose among matching cameras by their validated positive relative weights.
+    total = sum(weight for weight, _ in renderable)
+    digest = hashlib.blake2b(f"vqa:{sample_idx}".encode(), digest_size=8).digest()
+    draw = int.from_bytes(digest, "big") / 2**64 * total
+    cumulative = 0.0
+    for weight, rendered in renderable:
+        cumulative += weight
+        if draw < cumulative:
+            return rendered
+    return renderable[-1][1]
+
+
 def _select_recipe(recipe: TrainingRecipe, sample_idx: int) -> TrainingRecipe:
    """Pick a deterministic blend component for ``sample_idx`` (or return ``recipe``)."""
    if recipe.blend is None:
@@ -201,7 +271,8 @@ def _select_recipe(recipe: TrainingRecipe, sample_idx: int) -> TrainingRecipe:
        cumulative += component.weight or 0.0
        if draw < cumulative:
            return component
-    assert last_component is not None
+    if last_component is None:
+        raise ValueError("Blend recipes must contain at least one component.")
    return last_component


@@ -321,7 +392,8 @@ def _render_message_recipe(
    bindings: dict[str, LanguageRow | str | None],
 ) -> RenderedMessages | None:
    """Expand ``recipe.messages`` into rendered chat messages using ``bindings``."""
-    assert recipe.messages is not None
+    if recipe.messages is None:
+        raise ValueError("Cannot render a blend recipe as a message recipe.")
    messages: list[dict[str, Any]] = []
    streams: list[str | None] = []
    target_indices: list[int] = []
@@ -346,7 +418,9 @@ def _render_message_recipe(
        if turn.target:
            target_indices.append(message_idx)

-    if not target_indices:
+    # Keep samples with either text targets or low-level action supervision.
+    has_low_level = any(stream == "low_level" for stream in streams)
+    if not target_indices and not has_low_level:
        return None

    rendered = {
@@ -403,14 +477,12 @@ def _validate_rendered(rendered: RenderedMessages) -> None:

    if len(streams) != len(messages):
        raise ValueError("message_streams must be aligned with messages.")
-    if not target_indices:
-        raise ValueError("Rendered samples must contain at least one target message.")
+    # Require text or low-level action supervision.
+    if not target_indices and not any(s == "low_level" for s in streams):
+        raise ValueError("Rendered samples must contain a target message or a low_level-stream message.")
    for idx in target_indices:
        if idx < 0 or idx >= len(messages):
            raise ValueError(f"Target message index {idx} is out of bounds.")
-    # ``stream`` is enforced non-None at MessageTurn construction time
-    # (see ``MessageTurn.__post_init__``), so a missing stream here would
-    # mean the dataclass invariant was bypassed; no need to re-check.


 def _nth_relative(
@@ -18,6 +18,7 @@ import dataclasses
 import importlib.resources
 import json
 import logging
+from collections.abc import Sequence
 from dataclasses import dataclass, field
 from pathlib import Path

@@ -98,6 +99,47 @@ VIDEO_DIR = "videos"

 CHUNK_FILE_PATTERN = "chunk-{chunk_index:03d}/file-{file_index:03d}"
 IMAGE_FILE_PATTERN = "frame-{frame_index:06d}.png"
+
+
+def resolve_episode_indices(
+    episodes: Sequence[int] | None,
+    total_episodes: int,
+    exclude_episodes: Sequence[int] | None = None,
+) -> list[int] | None:
+    """Resolve an optional episode allowlist and exclusion list against dataset bounds.
+
+    ``None`` is preserved when no filtering is requested so callers can retain
+    their native "all episodes" fast path. Invalid indices are ignored with a
+    warning, and the input order is preserved.
+    """
+    if total_episodes < 0:
+        raise ValueError(f"total_episodes must be non-negative, got {total_episodes}")
+
+    if episodes is None and not exclude_episodes:
+        return None
+
+    candidates = list(range(total_episodes)) if episodes is None else list(episodes)
+    invalid = [episode for episode in candidates if not 0 <= episode < total_episodes]
+    if invalid:
+        logger.warning(
+            "Ignoring episode indices outside the dataset range [0, %d): %s",
+            total_episodes,
+            invalid,
+        )
+    candidates = [episode for episode in candidates if 0 <= episode < total_episodes]
+
+    excluded = set(exclude_episodes or [])
+    invalid_excluded = sorted(episode for episode in excluded if not 0 <= episode < total_episodes)
+    if invalid_excluded:
+        logger.warning(
+            "Ignoring excluded episode indices outside the dataset range [0, %d): %s",
+            total_episodes,
+            invalid_excluded,
+        )
+    excluded = {episode for episode in excluded if 0 <= episode < total_episodes}
+    return [episode for episode in candidates if episode not in excluded]
+
+
 DEPTH_FILE_PATTERN = "frame-{frame_index:06d}.tiff"
 DEFAULT_TASKS_PATH = "meta/tasks.parquet"
 DEFAULT_EPISODES_PATH = EPISODES_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
@@ -0,0 +1,43 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Distributed-training runtime for LeRobot.
+
+This package owns everything that turns the declarative topology in
+:class:`lerobot.configs.parallelism.ParallelismConfig` into a running engine:
+mesh math (:class:`~lerobot.distributed.parallel_dims.ParallelDims`), the
+`Accelerator` factory (:func:`~lerobot.distributed.factory.make_accelerator`),
+sharding-aware checkpoint helpers, and small rank utilities.
+
+Setup-order contract (normative):
+CP dispatch install -> activation checkpointing -> torch.compile ->
+``fully_shard``/DDP (via ``accelerator.prepare``) -> optimizer rebind.
+Only the last two steps are active today; CP/AC/compile are configured
+placeholders wired in later rounds.
+"""
+
+from .factory import guard_against_env_interference, make_accelerator, set_fsdp_wrap_modules
+from .parallel_dims import ParallelDims
+from .utils import finalize_sharded_policy, is_main_process, strip_accelerate_cp_hooks
+
+__all__ = [
+    "ParallelDims",
+    "finalize_sharded_policy",
+    "guard_against_env_interference",
+    "is_main_process",
+    "make_accelerator",
+    "set_fsdp_wrap_modules",
+    "strip_accelerate_cp_hooks",
+]
@@ -0,0 +1,195 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Sharding-aware checkpoint primitives.
+
+Two artifact channels with distinct owners:
+
+- the **distributable** ``model.safetensors``: produced by ``PreTrainedPolicy.save_pretrained``
+  through :func:`full_model_state_dict` — a collective full gather when the model is sharded;
+- the **resume** channel (sharded runs): torch DCP directories written/read through accelerate's
+  ``save/load_fsdp_model`` and ``save/load_fsdp_optimizer`` (``pytorch_model_fsdp_0/`` and
+  ``optimizer_0/``, names imported from accelerate constants), which reshard on load across
+  topology changes.
+
+Every function that touches sharded state is a collective and must run on ALL ranks.
+"""
+
+from pathlib import Path
+from typing import TYPE_CHECKING
+
+import torch
+from torch import nn
+
+if TYPE_CHECKING:
+    from accelerate import Accelerator
+
+
+def is_sharded_module(module: nn.Module) -> bool:
+    """True when `fully_shard` owns this module's parameters (FSDP2's in-place class swap).
+
+    Args:
+        module (nn.Module): The module to inspect (a torch.compile wrapper is looked through
+            via `_orig_mod`).
+
+    Returns:
+        bool: True when the module (or its compiled `_orig_mod`) is an `FSDPModule`.
+    """
+    from torch.distributed.fsdp import FSDPModule
+
+    if isinstance(module, FSDPModule):
+        return True
+    # torch.compile wraps the sharded module; mirror accelerate's `_orig_mod` check.
+    orig_mod = getattr(module, "_orig_mod", None)
+    return orig_mod is not None and isinstance(orig_mod, FSDPModule)
+
+
+def full_model_state_dict(module: nn.Module) -> dict[str, torch.Tensor]:
+    """The module's full (unsharded) state dict, however its parameters are laid out.
+
+    Sharded modules gather through torch's DCP state-dict API: a COLLECTIVE that must run on
+    every rank; with ``cpu_offload=True`` the full dict materializes on the main rank only and
+    every other rank receives a literal ``{}`` (runtime-verified — a
+    rank-0-gated call deadlocks). Plain modules return ``module.state_dict()`` on every rank.
+
+    Args:
+        module (nn.Module): The (possibly sharded) module to read the state dict from.
+
+    Returns:
+        dict[str, torch.Tensor]: The full state dict — on the main rank only (``{}``
+            elsewhere) when the module is sharded, on every rank otherwise.
+    """
+    if not is_sharded_module(module):
+        return module.state_dict()
+
+    from torch.distributed.checkpoint.state_dict import StateDictOptions, get_model_state_dict
+
+    return get_model_state_dict(module, options=StateDictOptions(full_state_dict=True, cpu_offload=True))
+
+
+def _fsdp_plugin(accelerator: "Accelerator") -> object:
+    """The accelerator's FSDP plugin, required by every DCP save/load helper below.
+
+    Args:
+        accelerator (Accelerator): The accelerator that prepared the sharded model.
+
+    Returns:
+        object: The FSDP plugin held by `accelerator.state`.
+
+    Raises:
+        RuntimeError: If the accelerator was not configured with an FSDP plugin.
+    """
+    plugin = getattr(accelerator.state, "fsdp_plugin", None)
+    if plugin is None:
+        raise RuntimeError("Sharded checkpointing requires an FSDP-prepared Accelerator.")
+    return plugin
+
+
+def save_sharded_model(accelerator: "Accelerator", model: nn.Module, output_dir: Path) -> None:
+    """Write the DCP model shards (`pytorch_model_fsdp_0/`). Collective: call on all ranks.
+
+    Args:
+        accelerator (Accelerator): The accelerator that prepared the sharded model.
+        model (nn.Module): The prepared (sharded) model to save.
+        output_dir (Path): The directory the shard subdirectory is created in.
+    """
+    from accelerate.utils import save_fsdp_model
+
+    # accelerate 1.14's DCP helpers do string containment checks on the path:
+    # always hand them str, never Path.
+    save_fsdp_model(_fsdp_plugin(accelerator), accelerator, model, str(output_dir))
+
+
+def load_sharded_model(accelerator: "Accelerator", model: nn.Module, input_dir: Path) -> None:
+    """Load DCP model shards into the prepared (sharded) model. Collective: call on all ranks.
+
+    Args:
+        accelerator (Accelerator): The accelerator that prepared the sharded model.
+        model (nn.Module): The prepared (sharded) model to load into.
+        input_dir (Path): The directory containing the `pytorch_model_fsdp_0/` shard
+            subdirectory.
+    """
+    from accelerate.utils import load_fsdp_model
+    from accelerate.utils.constants import FSDP_MODEL_NAME
+
+    # Pass the exact shard directory: accelerate's load resolves it with a substring check
+    # ("pytorch_model_fsdp" in the path -> use as-is), which misfires on run paths that happen
+    # to contain the marker; the exact dir makes the check deterministic.
+    load_fsdp_model(_fsdp_plugin(accelerator), accelerator, model, str(input_dir / f"{FSDP_MODEL_NAME}_0"))
+
+
+def save_sharded_optimizer(
+    accelerator: "Accelerator", optimizer: torch.optim.Optimizer, model: nn.Module, output_dir: Path
+) -> None:
+    """Write the DCP optimizer shards (`optimizer_0/`). Collective: call on all ranks.
+
+    Args:
+        accelerator (Accelerator): The accelerator that prepared the model and optimizer.
+        optimizer (torch.optim.Optimizer): The prepared optimizer to save the state from.
+        model (nn.Module): The prepared (sharded) model the optimizer state is keyed by.
+        output_dir (Path): The directory the shard subdirectory is created in.
+    """
+    from accelerate.utils import save_fsdp_optimizer
+
+    save_fsdp_optimizer(_fsdp_plugin(accelerator), accelerator, optimizer, model, str(output_dir))
+
+
+def load_sharded_optimizer(
+    accelerator: "Accelerator", optimizer: torch.optim.Optimizer, model: nn.Module, input_dir: Path
+) -> None:
+    """Load DCP optimizer shards into the prepared optimizer. Collective: call on all ranks.
+
+    Must run AFTER ``accelerator.prepare()``: FSDP2's prepare rebinds the optimizer's param
+    groups to sharded DTensors but never migrates ``optimizer.state`` — the resharding load is
+    the only correct way to restore it.
+
+    Args:
+        accelerator (Accelerator): The accelerator that prepared the model and optimizer.
+        optimizer (torch.optim.Optimizer): The prepared optimizer to restore the state into.
+        model (nn.Module): The prepared (sharded) model the optimizer state is keyed by.
+        input_dir (Path): The directory containing the `optimizer_0/` shard subdirectory.
+    """
+    from accelerate.utils import load_fsdp_optimizer
+    from accelerate.utils.constants import OPTIMIZER_NAME
+
+    # Exact shard directory for the same reason as load_sharded_model: accelerate's substring
+    # check ("optimizer" in the path) would misread e.g. --job_name=optimizer_sweep run paths.
+    load_fsdp_optimizer(
+        _fsdp_plugin(accelerator), accelerator, optimizer, model, str(input_dir / f"{OPTIMIZER_NAME}_0")
+    )
+
+
+def dcp_to_safetensors(dcp_dir: Path, output_dir: Path, *, delete_dcp: bool = False) -> Path:
+    """Merge a DCP shard directory into a single `model.safetensors` (offline, single process).
+
+    Thin wrapper over `accelerate.utils.merge_fsdp_weights`, which loads the shards without a
+    process group, writes safetensors directly, and — when asked — removes the merged shard
+    directory itself, only on the main process and only once the merge has succeeded.
+
+    Args:
+        dcp_dir (Path): The DCP shard directory to merge (e.g. `.../pytorch_model_fsdp_0`).
+        output_dir (Path): The directory the merged `model.safetensors` is written into.
+        delete_dcp (bool): Whether to remove the shard directory once it has been merged.
+            Defaults to False.
+
+    Returns:
+        Path: The written `model.safetensors` file's path.
+    """
+    from accelerate.utils import merge_fsdp_weights
+
+    merge_fsdp_weights(
+        str(dcp_dir), str(output_dir), safe_serialization=True, remove_checkpoint_dir=delete_dcp
+    )
+    return output_dir / "model.safetensors"
@@ -0,0 +1,139 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""The `Accelerator` factory — the only place accelerate gets configured.
+
+`torchrun` is the launcher; every accelerate parameter comes from `TrainPipelineConfig`
+(`cfg.parallelism` + `cfg.accelerator`) so a run is reproducible from its `train_config.json`
+alone. `accelerate launch` without a `--config_file` remains equivalent (it only sets rendezvous
+env vars in that mode); the yaml flow is superseded.
+"""
+
+import os
+from typing import TYPE_CHECKING
+
+from lerobot.configs.parallelism import world_size_from_env
+from lerobot.configs.train import TrainPipelineConfig
+
+if TYPE_CHECKING:
+    from accelerate import Accelerator
+
+    from lerobot.policies.pretrained import PreTrainedPolicy
+
+# Env vars through which `accelerate launch --config_file` (or a stray shell) would configure
+# accelerate behind the config system's back, making train_config.json lie about what ran.
+_ACCELERATE_ENV_VARS = (
+    "ACCELERATE_USE_FSDP",
+    "ACCELERATE_USE_PARALLELISM_CONFIG",
+    "ACCELERATE_GRADIENT_ACCUMULATION_STEPS",
+)
+_ENV_OVERRIDE = "LEROBOT_ALLOW_ACCELERATE_ENV"
+
+
+def guard_against_env_interference() -> None:
+    """Hard-error when accelerate-configuring env vars are set.
+
+    A silently env-overridden "reproducible" config is worse than a stop: users migrating from
+    the old `accelerate launch --config_file fsdp.yaml` flow get a precise error instead of a
+    config that lies. Set LEROBOT_ALLOW_ACCELERATE_ENV=1 to acknowledge and proceed.
+
+    Raises:
+        RuntimeError: If any accelerate-configuring environment variable is set and the
+            LEROBOT_ALLOW_ACCELERATE_ENV override is not.
+    """
+    if os.environ.get(_ENV_OVERRIDE):
+        return
+    offending = sorted(name for name in _ACCELERATE_ENV_VARS if name in os.environ)
+    if offending:
+        raise RuntimeError(
+            f"Accelerate-configuring environment variables are set: {', '.join(offending)}. "
+            "LeRobot manages accelerate exclusively through TrainPipelineConfig "
+            "(--parallelism.* / --accelerator.*); launch with plain torchrun and remove these "
+            "variables (the `accelerate launch --config_file` flow is superseded), or set "
+            f"{_ENV_OVERRIDE}=1 to acknowledge that they may override your config."
+        )
+
+
+def make_accelerator(cfg: TrainPipelineConfig) -> "Accelerator":
+    """Resolve the topology against the launched world and build the `Accelerator`.
+
+    Must run once per process, before any other component needs the device or the process
+    group (`Accelerator.__init__` initializes both and builds the device mesh).
+
+    Args:
+        cfg (TrainPipelineConfig): The full training config; `cfg.parallelism` is resolved in
+            place against the launched world size and `cfg.accelerator` builds the result.
+
+    Returns:
+        Accelerator: The configured accelerator, with device and process group initialized.
+
+    Raises:
+        ValueError: If `cfg.checkpoint_format` requires DCP but the topology resolved to a
+            non-sharded run.
+    """
+    guard_against_env_interference()
+    cfg.parallelism.resolve(world_size_from_env())
+    # The parse-time format check ran against the declared degrees, where the dp_shard=-1
+    # sentinel counts as sharded; it may resolve to an unsharded run (e.g. -1 at world size 1).
+    # Re-check against the concrete degrees so the recorded format never lies about the
+    # artifacts a checkpoint will actually contain.
+    if cfg.checkpoint_format.wants_dcp and not cfg.parallelism.is_sharded:
+        raise ValueError(
+            f"checkpoint_format={cfg.checkpoint_format.value} requires a sharded run, but the "
+            f"topology resolved to a non-sharded one (dp_replicate={cfg.parallelism.dp_replicate}, "
+            f"dp_shard={cfg.parallelism.dp_shard}); non-sharded checkpoints are always safetensors."
+        )
+    return cfg.accelerator.build(
+        cfg.parallelism,
+        cpu=cfg.trainable_config.device == "cpu",
+    )
+
+
+def set_fsdp_wrap_modules(accelerator: "Accelerator", policy: "PreTrainedPolicy") -> None:
+    """Resolve the FSDP wrap-unit class names onto the plugin before `accelerator.prepare()`.
+
+    Resolution order: user override (`--accelerator.fsdp.wrap_modules`, already on the plugin)
+    -> the policy's `_fsdp_wrap_modules` declaration -> hard error. Root-only wrapping — the
+    silent default when no wrap source exists — is never accepted: it quietly forfeits all
+    sharding memory savings.
+
+    No-op for the size-based policy (`--accelerator.fsdp.min_num_params`), which needs no class
+    names, and for non-sharded runs (no fsdp plugin).
+
+    Args:
+        accelerator (Accelerator): The accelerator whose FSDP plugin receives the wrap-unit
+            class names.
+        policy (PreTrainedPolicy): The trainable whose class may declare `_fsdp_wrap_modules`.
+
+    Raises:
+        ValueError: If sharded class-based wrapping is configured but neither a user override
+            nor a policy declaration supplies wrap-unit class names.
+    """
+    plugin = getattr(accelerator.state, "fsdp_plugin", None)
+    if plugin is None or plugin.min_num_params:
+        return
+    if plugin.transformer_cls_names_to_wrap:  # user override, set at build time
+        return
+    # getattr, not attribute access: non-policy trainables (no `_fsdp_wrap_modules` attribute)
+    # must reach the actionable error below, not an AttributeError.
+    declared = getattr(type(policy), "_fsdp_wrap_modules", None)
+    if not declared:
+        raise ValueError(
+            f"Policy '{type(policy).__name__}' declares no FSDP wrap units. Sharded training "
+            "requires wrap-unit class names: set --accelerator.fsdp.wrap_modules='[\"MyBlock\"]' "
+            "(or --accelerator.fsdp.min_num_params for a size-based policy), or declare "
+            "`_fsdp_wrap_modules` on the policy class."
+        )
+    plugin.transformer_cls_names_to_wrap = list(declared)
@@ -0,0 +1,112 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Runtime mesh math derived from the declarative :class:`ParallelismConfig`.
+
+`ParallelDims` is the training script's single source of truth for topology-derived numbers
+(data-parallel world size and rank, sample accounting inputs) and — once the CP engine lands —
+the owner of LeRobot's private ``(dp_replicate, dp_shard, ring, ulysses)`` mesh. It is a runtime
+object and is never serialized (the config it derives from is what lands in
+``train_config.json``).
+"""
+
+from dataclasses import dataclass
+
+import torch.distributed as dist
+
+from lerobot.configs.parallelism import ParallelismConfig
+
+
+@dataclass(frozen=True)
+class ParallelDims:
+    """Concrete parallelism degrees bound to a world size (canonical row-major rank layout)."""
+
+    dp_replicate: int
+    dp_shard: int
+    ring: int
+    ulysses: int
+    world_size: int
+    device_type: str
+
+    @classmethod
+    def from_config(cls, cfg: ParallelismConfig, world_size: int, device_type: str) -> "ParallelDims":
+        """Bind a *resolved* config to the actual runtime world size (cross-checked here).
+
+        Args:
+            cfg (ParallelismConfig): The declarative topology, already resolved via
+                `ParallelismConfig.resolve(world_size)`.
+            world_size (int): The launched world size the declared degrees must multiply to.
+            device_type (str): The accelerator device type backing the mesh (e.g. "cuda").
+
+        Returns:
+            ParallelDims: The concrete parallelism degrees bound to this world.
+
+        Raises:
+            ValueError: If the config is unresolved (`dp_shard == -1`) or its degrees do not
+                multiply to `world_size`.
+        """
+        total = cfg.dp_replicate * cfg.dp_shard * cfg.cp_size
+        if cfg.dp_shard == -1 or total != world_size:
+            raise ValueError(
+                f"ParallelismConfig is not resolved against this world: dp_replicate="
+                f"{cfg.dp_replicate} * dp_shard={cfg.dp_shard} * cp={cfg.cp_size} != "
+                f"world_size={world_size}. Call ParallelismConfig.resolve(world_size) first "
+                "(make_accelerator does this)."
+            )
+        return cls(
+            dp_replicate=cfg.dp_replicate,
+            dp_shard=cfg.dp_shard,
+            ring=cfg.context_parallel.ring_degree,
+            ulysses=cfg.context_parallel.ulysses_degree,
+            world_size=world_size,
+            device_type=device_type,
+        )
+
+    @property
+    def cp_size(self) -> int:
+        """Total context-parallel degree (`ring * ulysses`)."""
+        return self.ring * self.ulysses
+
+    @property
+    def is_sharded(self) -> bool:
+        """Whether parameters are sharded (`dp_shard > 1` or any context parallelism)."""
+        return self.dp_shard > 1 or self.cp_size > 1
+
+    @property
+    def dp_world_size(self) -> int:
+        """Number of distinct data-parallel workers — the divisor for all sample accounting."""
+        return self.dp_replicate * self.dp_shard
+
+    @property
+    def dp_rank(self) -> int:
+        """This process's data-parallel coordinate (CP peers share one dp_rank).
+
+        With the canonical row-major layout and (ring, ulysses) innermost, CP peers are
+        contiguous global ranks, so the dp coordinate is the integer quotient by cp_size —
+        the same arithmetic accelerate's mesh-aware dataloader applies.
+        """
+        global_rank = dist.get_rank() if dist.is_initialized() else 0
+        return global_rank // self.cp_size
+
+    def cp_mesh(self) -> None:
+        """Private (ring, ulysses) mesh for the CP engine — reserved for the CP round.
+
+        Raises:
+            NotImplementedError: Always — context parallelism is not implemented yet.
+        """
+        raise NotImplementedError(
+            "Context parallelism is not implemented yet; ParallelDims.cp_mesh is reserved for "
+            "the CP engine round (a private mesh aligned with accelerate's cp block)."
+        )
@@ -0,0 +1,94 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Rank utilities and post-`prepare()` sharding finalization."""
+
+import logging
+from typing import TYPE_CHECKING
+
+import torch.distributed as dist
+from torch import nn
+
+if TYPE_CHECKING:
+    from lerobot.distributed.parallel_dims import ParallelDims
+
+
+def is_main_process() -> bool:
+    """True on the process that owns rank-0-only side effects (file writes, uploads, logging).
+
+    Torch-native on purpose: persistence code must not depend on an `Accelerator` handle —
+    `_save_pretrained` and the hub publishers run in contexts that have none. Outside
+    distributed runs every process is the main process.
+
+    Returns:
+        bool: True when this process is rank 0 or no process group is initialized.
+    """
+    return not dist.is_initialized() or dist.get_rank() == 0
+
+
+def strip_accelerate_cp_hooks(model: nn.Module) -> int:
+    """Remove accelerate's context-parallel forward-pre-hooks from every module.
+
+    When `cp_size > 1` is declared, `accelerator.prepare()` unconditionally attaches hooks that
+    silently replace any `attention_mask` kwarg of `*self_attn` modules with `is_causal=True`
+    (`accelerate.big_modeling._attach_context_parallel_hooks`) — mask corruption for policies
+    with non-causal attention. LeRobot implements CP itself and never enters accelerate's CP
+    context, so these hooks are pure hazard. Deterministically identified by their defining
+    module; a version canary pins that identity.
+
+    Args:
+        model (nn.Module): The prepared model to strip the hooks from (all submodules are
+            visited).
+
+    Returns:
+        int: The number of hooks removed.
+    """
+    removed = 0
+    for module in model.modules():
+        for hook_id, hook in list(module._forward_pre_hooks.items()):
+            if getattr(hook, "__module__", None) == "accelerate.big_modeling":
+                del module._forward_pre_hooks[hook_id]
+                module._forward_pre_hooks_with_kwargs.pop(hook_id, None)
+                removed += 1
+    return removed
+
+
+def finalize_sharded_policy(policy: nn.Module, parallel_dims: "ParallelDims") -> None:
+    """Sharding correctness protocol, applied once, immediately after `accelerator.prepare()`.
+
+    1. Strip accelerate's CP mask hooks (only attached when cp > 1 was declared).
+    2. Register the policy's non-`forward` entry points (`_fsdp_forward_methods`) so FSDP2
+       unshards parameters around `select_action` & co. — without this, any inference-style
+       call on a sharded policy crashes on mixed Tensor/DTensor.
+
+    No-op for DDP/single-process runs.
+
+    Args:
+        policy (nn.Module): The policy as returned by `accelerator.prepare()`.
+        parallel_dims (ParallelDims): The run's resolved topology; decides whether the protocol
+            applies.
+    """
+    if not parallel_dims.is_sharded:
+        return
+    if parallel_dims.cp_size > 1:
+        removed = strip_accelerate_cp_hooks(policy)
+        logging.info("Stripped %d accelerate context-parallel attention-mask hooks.", removed)
+
+    from torch.distributed.fsdp import FSDPModule, register_fsdp_forward_method
+
+    if isinstance(policy, FSDPModule):
+        for method_name in getattr(type(policy), "_fsdp_forward_methods", ()):
+            if callable(getattr(policy, method_name, None)):
+                register_fsdp_forward_method(policy, method_name)
@@ -328,6 +328,7 @@ class LiberoEnv(EnvConfig):
    render_mode: str = "rgb_array"
    camera_name: str = "agentview_image,robot0_eye_in_hand_image"
    init_states: bool = True
+    hard_reset: bool = True
    camera_name_mapping: dict[str, str] | None = None
    observation_height: int = 360
    observation_width: int = 360
@@ -356,6 +357,8 @@ class LiberoEnv(EnvConfig):
    def __post_init__(self):
        if self.fps <= 0:
            raise ValueError(f"fps must be positive, got {self.fps}")
+        if not self.hard_reset and not self.init_states:
+            raise ValueError("hard_reset=False requires init_states=True")

        if self.obs_type == "pixels":
            self.features[LIBERO_KEY_PIXELS_AGENTVIEW] = PolicyFeature(
@@ -416,6 +419,7 @@ class LiberoEnv(EnvConfig):
            "observation_height": self.observation_height,
            "observation_width": self.observation_width,
            "control_freq": self.fps,
+            "hard_reset": self.hard_reset,
        }
        if self.task_ids is not None:
            kwargs["task_ids"] = self.task_ids
@@ -128,10 +128,13 @@ class LiberoEnv(gym.Env):
        control_freq: int = 20,
        control_mode: str = "relative",
        is_libero_plus: bool = False,
+        hard_reset: bool = True,
    ):
        super().__init__()
        if control_freq <= 0:
            raise ValueError(f"control_freq must be positive, got {control_freq}")
+        if not hard_reset and not init_states:
+            raise ValueError("hard_reset=False requires init_states=True")
        self.task_id = task_id
        self.is_libero_plus = is_libero_plus
        self.obs_type = obs_type
@@ -158,6 +161,7 @@ class LiberoEnv(gym.Env):
        self.camera_name_mapping = camera_name_mapping
        self.num_steps_wait = num_steps_wait
        self.control_freq = control_freq
+        self.hard_reset = hard_reset
        self.episode_index = episode_index
        self.episode_length = episode_length
        # Load once and keep
@@ -265,6 +269,9 @@ class LiberoEnv(gym.Env):
            camera_heights=self.observation_height,
            camera_widths=self.observation_width,
            control_freq=self.control_freq,
+            # Soft resets skip LIBERO's model and renderer rebuild. They are opt-in
+            # because settle steps can make their observations differ from hard resets.
+            hard_reset=self.hard_reset,
        )
        env.reset()
        self._env = env
@@ -377,8 +384,9 @@ class LiberoEnv(gym.Env):
            }
        )
        observation = self._format_raw_obs(raw_obs)
-        if terminated:
-            self.reset()
+        # Return the terminal observation unchanged. The caller owns resetting after
+        # termination; vector envs created below use NEXT_STEP autoreset. Resetting here
+        # would therefore reset twice and skip an initial state.
        truncated = False
        return observation, reward, terminated, truncated, info

@@ -476,6 +484,7 @@ def create_libero_envs(
        print(f"Restricting to task_ids={task_ids_filter}")

    is_async = env_cls is gym.vector.AsyncVectorEnv
+    is_sync = env_cls is gym.vector.SyncVectorEnv

    out: dict[str, dict[int, Any]] = defaultdict(dict)
    for suite_name in suite_names:
@@ -512,6 +521,10 @@ def create_libero_envs(
                    cached_act_space = lazy.action_space
                    cached_metadata = lazy.metadata
                out[suite_name][tid] = lazy
+            elif is_sync:
+                out[suite_name][tid] = gym.vector.SyncVectorEnv(
+                    fns, autoreset_mode=gym.vector.AutoresetMode.NEXT_STEP
+                )
            else:
                out[suite_name][tid] = env_cls(fns)
            print(f"Built vec env | suite={suite_name} | task_id={tid} | n_envs={n_envs}")
@@ -177,6 +177,76 @@ def _sub_env_has_attr(env: gym.vector.VectorEnv, attr: str) -> bool:
        return False


+# Passed in `reset(options=...)` by `rollout()` to mark the start of a new rollout.
+# FreezeAfterEpisodeEnd thaws only on this, so Gymnasium's argument-less autoreset
+# cannot be mistaken for a genuine new episode.
+NEW_ROLLOUT_OPTION = "lerobot_new_rollout"
+
+
+class FreezeAfterEpisodeEnd(gym.Wrapper):
+    """Stop doing simulator work once a sub-env's episode has ended.
+
+    `rollout()` runs `while not np.all(done)` with `done` latched, so a sub-env that
+    terminates early keeps being stepped -- physics and offscreen rendering included --
+    until the slowest sub-env in the batch finishes. The batch runs for
+    `max(episode_lengths)` iterations to complete work that only needs
+    `mean(episode_lengths)`.
+
+    This caches the terminal transition and replays it for any further `step()` or
+    autoreset, so a finished sub-env costs nothing. The rollout already ignores those
+    transitions.
+
+    The freeze survives Gymnasium's autoreset deliberately. Under
+    `AutoresetMode.NEXT_STEP` the vector env resets a terminated sub-env on the
+    following step and runs it through an entire extra episode that the rollout
+    discards, because `done` stays latched. Absorbing that reset is most of the saving.
+
+    Only an explicit reset carrying `NEW_ROLLOUT_OPTION` thaws it, so the signal is
+    explicit rather than inferred: Gymnasium's autoreset calls `reset()` with no
+    arguments, but so would a caller passing `seeds=None`, and confusing the two would
+    strand an env frozen for a whole rollout.
+
+    `AutoresetMode.DISABLED` is not an alternative here — Gymnasium asserts that no
+    terminated env is ever stepped in that mode, so the wrapper is never reached.
+    """
+
+    def __init__(self, env: gym.Env):
+        super().__init__(env)
+        self._frozen: tuple | None = None
+
+    def reset(self, *, seed=None, options=None):
+        if self._frozen is not None and not (options or {}).get(NEW_ROLLOUT_OPTION):
+            # Gymnasium's autoreset for a sub-env the rollout has already finished with.
+            # Replay the terminal observation instead of rebuilding the simulation.
+            obs, _, _, _, info = self._frozen
+            return obs, info
+        self._frozen = None
+        return self.env.reset(seed=seed, options=options)
+
+    def step(self, action):
+        if self._frozen is not None:
+            return self._frozen
+        obs, reward, terminated, truncated, info = self.env.step(action)
+        if terminated or truncated:
+            # Zero the reward on replay so a frozen sub-env cannot inflate a return if a
+            # caller sums rewards over the padded tail.
+            self._frozen = (obs, 0.0, terminated, truncated, info)
+        return obs, reward, terminated, truncated, info
+
+    @property
+    def is_frozen(self) -> bool:
+        return self._frozen is not None
+
+
+def freeze_after_episode_end(env_fn: Callable[[], gym.Env]) -> Callable[[], gym.Env]:
+    """Wrap an env factory so the built env freezes once its episode ends."""
+
+    def _fn() -> gym.Env:
+        return FreezeAfterEpisodeEnd(env_fn())
+
+    return _fn
+
+
 class _LazyAsyncVectorEnv:
    """Defers AsyncVectorEnv creation until first use.

@@ -212,7 +282,12 @@ class _LazyAsyncVectorEnv:

    def _ensure(self) -> None:
        if self._env is None:
-            self._env = gym.vector.AsyncVectorEnv(self._env_fns, context="forkserver", shared_memory=True)
+            self._env = gym.vector.AsyncVectorEnv(
+                [freeze_after_episode_end(fn) for fn in self._env_fns],
+                context="forkserver",
+                shared_memory=True,
+                autoreset_mode=gym.vector.AutoresetMode.NEXT_STEP,
+            )

    @property
    def unwrapped(self):
@@ -432,7 +432,7 @@ def submit_to_hf(cfg: TrainPipelineConfig) -> None:

    # Finish as soon as the model is pushed, rather than waiting out the platform's
    # post-run finalization before the job stage flips to COMPLETED. This matches the
-    # exact log line emitted by PreTrainedPolicy.push_model_to_hub — the two must stay
+    # exact log line emitted by lerobot.common.train_utils.publish_trained_model — the two must stay
    # in sync. If it ever stops matching we just fall back to stage-based completion
    # (~30s slower), so the contract is an optimization, not a correctness requirement.
    success_marker = f"Model pushed to https://huggingface.co/{repo_id}"
@@ -20,7 +20,6 @@ from .optimizers import (
    SGDConfig as SGDConfig,
    XVLAAdamWConfig as XVLAAdamWConfig,
    load_optimizer_state,
-    load_optimizer_state_dict,
    save_optimizer_state,
 )
 from .schedulers import (
@@ -51,7 +50,6 @@ __all__ = [
    "VQBeTSchedulerConfig",
    # State management
    "load_optimizer_state",
-    "load_optimizer_state_dict",
    "load_scheduler_state",
    "save_optimizer_state",
    "save_scheduler_state",
@@ -27,7 +27,7 @@ from lerobot.utils.constants import (
    OPTIMIZER_PARAM_GROUPS,
    OPTIMIZER_STATE,
 )
-from lerobot.utils.io_utils import deserialize_json_into_object, load_json, write_json
+from lerobot.utils.io_utils import deserialize_json_into_object, write_json
 from lerobot.utils.utils import flatten_dict, unflatten_dict

 # Type alias for parameters accepted by optimizer build() methods.
@@ -52,6 +52,11 @@ class OptimizerConfig(draccus.ChoiceRegistry, abc.ABC):
    def type(self) -> str:
        return self.get_choice_name(self.__class__)

+    @property
+    def builds_multiple_optimizers(self) -> bool:
+        """True when build() returns a dict of optimizers (unsupported under sharded training)."""
+        return False
+
    @classmethod
    def default_choice_name(cls) -> str | None:
        return "adam"
@@ -245,6 +250,10 @@ class MultiAdamConfig(OptimizerConfig):
    grad_clip_norm: float = 10.0
    optimizer_groups: dict[str, dict[str, Any]] = field(default_factory=dict)

+    @property
+    def builds_multiple_optimizers(self) -> bool:
+        return True
+
    def build(self, params: OptimizerParams) -> dict[str, torch.optim.Optimizer]:
        """Build multiple Adam optimizers.

@@ -283,35 +292,27 @@ class MultiAdamConfig(OptimizerConfig):
 def save_optimizer_state(
    optimizer: torch.optim.Optimizer | dict[str, torch.optim.Optimizer],
    save_dir: Path,
-    optim_state_dict: dict | None = None,
 ) -> None:
-    """Save optimizer state to disk.
+    """Save optimizer state to disk (non-sharded runs; sharded runs use the DCP channel).

    Args:
        optimizer: Either a single optimizer or a dictionary of optimizers.
        save_dir: Directory to save the optimizer state.
-        optim_state_dict: Pre-gathered optimizer state dict (for FSDP, where the sharded state must
-            be gathered across ranks first). If provided, it is saved directly instead of calling
-            ``optimizer.state_dict()``. Only supported for a single optimizer. Defaults to None.
    """
    if isinstance(optimizer, dict):
        # Handle dictionary of optimizers
-        if optim_state_dict is not None:
-            raise ValueError("optim_state_dict is not supported for a dict of optimizers")
        for name, opt in optimizer.items():
            optimizer_dir = save_dir / name
            optimizer_dir.mkdir(exist_ok=True, parents=True)
            _save_single_optimizer_state(opt, optimizer_dir)
    else:
        # Handle single optimizer
-        _save_single_optimizer_state(optimizer, save_dir, optim_state_dict=optim_state_dict)
+        _save_single_optimizer_state(optimizer, save_dir)


-def _save_single_optimizer_state(
-    optimizer: torch.optim.Optimizer, save_dir: Path, optim_state_dict: dict | None = None
-) -> None:
+def _save_single_optimizer_state(optimizer: torch.optim.Optimizer, save_dir: Path) -> None:
    """Save a single optimizer's state to disk."""
-    state = dict(optim_state_dict) if optim_state_dict is not None else optimizer.state_dict()
+    state = optimizer.state_dict()
    param_groups = state.pop("param_groups")
    flat_state = flatten_dict(state)
    save_file(flat_state, save_dir / OPTIMIZER_STATE)
@@ -365,19 +366,3 @@ def _load_single_optimizer_state(optimizer: torch.optim.Optimizer, save_dir: Pat

    optimizer.load_state_dict(loaded_state_dict)
    return optimizer
-
-
-def load_optimizer_state_dict(save_dir: Path) -> dict:
-    """Read a saved optimizer state dict (safetensors + json) back into a plain dict.
-
-    Unlike `load_optimizer_state`, this does not load into an optimizer and preserves the original
-    ``state`` keys verbatim (e.g. FSDP parameter FQNs, which are not integer-castable). It is used by
-    the FSDP resume path, where the full state must be resharded via `FSDP.optim_state_dict_to_load`
-    before being loaded into the (sharded) optimizer.
-    """
-    flat_state = load_file(save_dir / OPTIMIZER_STATE)
-    state = unflatten_dict(flat_state)
-    return {
-        "state": state.get("state", {}),
-        "param_groups": load_json(save_dir / OPTIMIZER_PARAM_GROUPS),
-    }
@@ -47,6 +47,8 @@ class ACTPolicy(PreTrainedPolicy):

    config_class = ACTConfig
    name = "act"
+    # FSDP2 wrap units: one unit per transformer layer of both stacks.
+    _fsdp_wrap_modules = ["ACTEncoderLayer", "ACTDecoderLayer"]

    def __init__(
        self,
@@ -31,10 +31,8 @@ from lerobot.envs import EnvConfig, env_to_policy_features
 from lerobot.lerobot_types import PolicyAction
 from lerobot.processor import (
    AbsoluteActionsProcessorStep,
-    NormalizerProcessorStep,
    PolicyProcessorPipeline,
    RelativeActionsProcessorStep,
-    UnnormalizerProcessorStep,
    batch_to_transition,
    policy_action_to_transition,
    transition_to_batch,
@@ -78,52 +76,6 @@ def _reconnect_relative_absolute_steps(
            step.relative_step = relative_step


-def _ensure_relative_actions(
-    preprocessor: PolicyProcessorPipeline, postprocessor: PolicyProcessorPipeline, policy_cfg
-) -> None:
-    """Enable (or inject) the relative/absolute action steps in a loaded pipeline.
-
-    When loading from a pretrained checkpoint, the saved processor is authoritative. If the base
-    predates the relative-action feature (e.g. FastWAM/LingBot bases) its pipeline has no
-    RelativeActionsProcessorStep, so lerobot-train's override cannot enable one — those override
-    keys are popped before `from_pretrained` (else it raises) and we reconstruct the steps here.
-    Bases that DO ship the (disabled) steps (e.g. pi0/pi05) are simply flipped on. No-op unless
-    ``policy_cfg.use_relative_actions`` is set, so non-relative runs are untouched.
-    """
-    if not getattr(policy_cfg, "use_relative_actions", False):
-        return
-
-    exclude_joints = list(getattr(policy_cfg, "relative_exclude_joints", []) or [])
-    action_names = getattr(policy_cfg, "action_feature_names", None)
-
-    pre_steps = list(preprocessor.steps)
-    relative_step = next((s for s in pre_steps if isinstance(s, RelativeActionsProcessorStep)), None)
-    if relative_step is None:
-        relative_step = RelativeActionsProcessorStep(
-            enabled=True, exclude_joints=exclude_joints, action_names=action_names
-        )
-        # Insert right before the normalizer (raw -> relative -> normalize); fall back to the front.
-        idx = next((i for i, s in enumerate(pre_steps) if isinstance(s, NormalizerProcessorStep)), 0)
-        pre_steps.insert(idx, relative_step)
-        preprocessor.steps = pre_steps
-    else:
-        relative_step.enabled = True
-        relative_step.exclude_joints = exclude_joints
-        relative_step.action_names = action_names
-
-    post_steps = list(postprocessor.steps)
-    absolute_step = next((s for s in post_steps if isinstance(s, AbsoluteActionsProcessorStep)), None)
-    if absolute_step is None:
-        absolute_step = AbsoluteActionsProcessorStep(enabled=True, relative_step=relative_step)
-        # Insert right after the unnormalizer (unnormalize -> absolute); fall back to the front.
-        idx = next((i for i, s in enumerate(post_steps) if isinstance(s, UnnormalizerProcessorStep)), -1)
-        post_steps.insert(idx + 1, absolute_step)
-        postprocessor.steps = post_steps
-    else:
-        absolute_step.enabled = True
-        absolute_step.relative_step = relative_step
-
-
 def get_policy_class(name: str) -> type[PreTrainedPolicy]:
    """
    Retrieves a policy class by its registered name.
@@ -245,20 +197,12 @@ def make_pre_post_processors(
                ),
            )

-        # The relative/absolute override keys only match if the saved base already contains those
-        # steps (e.g. pi0/pi05). For bases that predate the feature (FastWAM/LingBot) they would
-        # raise "Override keys ... do not match any step". Pop them here and let
-        # _ensure_relative_actions() enable-or-inject the steps after loading (handles both cases).
-        pre_overrides = dict(kwargs.get("preprocessor_overrides") or {})
-        post_overrides = dict(kwargs.get("postprocessor_overrides") or {})
-        pre_overrides.pop("relative_actions_processor", None)
-        post_overrides.pop("absolute_actions_processor", None)
        preprocessor = PolicyProcessorPipeline.from_pretrained(
            pretrained_model_name_or_path=pretrained_path,
            config_filename=kwargs.get(
                "preprocessor_config_filename", f"{POLICY_PREPROCESSOR_DEFAULT_NAME}.json"
            ),
-            overrides=pre_overrides,
+            overrides=kwargs.get("preprocessor_overrides", {}),
            to_transition=batch_to_transition,
            to_output=transition_to_batch,
            revision=pretrained_revision,
@@ -268,12 +212,11 @@ def make_pre_post_processors(
            config_filename=kwargs.get(
                "postprocessor_config_filename", f"{POLICY_POSTPROCESSOR_DEFAULT_NAME}.json"
            ),
-            overrides=post_overrides,
+            overrides=kwargs.get("postprocessor_overrides", {}),
            to_transition=policy_action_to_transition,
            to_output=transition_to_policy_action,
            revision=pretrained_revision,
        )
-        _ensure_relative_actions(preprocessor, postprocessor, policy_cfg)
        _reconnect_relative_absolute_steps(preprocessor, postprocessor)
        if isinstance(policy_cfg, Evo1Config):
            from .evo1.processor_evo1 import reconcile_evo1_processors
@@ -299,6 +242,7 @@ def make_policy(
    ds_meta: LeRobotDatasetMetadata | None = None,
    env_cfg: EnvConfig | None = None,
    rename_map: dict[str, str] | None = None,
+    defer_weight_load: bool = False,
 ) -> PreTrainedPolicy:
    """
    Instantiate a policy model.
@@ -309,22 +253,27 @@ def make_policy(
    can either initialize a new policy from scratch or load a pretrained one.

    Args:
-        cfg: The configuration for the policy to be created. If `cfg.pretrained_path` is
-             set, the policy will be loaded with weights from that path.
-        ds_meta: Dataset metadata used to infer feature shapes and types. Also provides
-                 statistics for normalization layers.
-        env_cfg: Environment configuration used to infer feature shapes and types.
-                 One of `ds_meta` or `env_cfg` must be provided.
-        rename_map: Optional mapping of dataset or environment feature keys to match
-                 expected policy feature names (e.g., `"left"` → `"camera1"`).
+        cfg (PreTrainedConfig): The configuration for the policy to be created. If
+            `cfg.pretrained_path` is set, the policy will be loaded with weights from that path.
+        ds_meta (LeRobotDatasetMetadata | None): Dataset metadata used to infer feature shapes and
+            types. Also provides statistics for normalization layers.
+        env_cfg (EnvConfig | None): Environment configuration used to infer feature shapes and
+            types. One of `ds_meta` or `env_cfg` must be provided.
+        rename_map (dict[str, str] | None): Optional mapping of dataset or environment feature
+            keys to match expected policy feature names (e.g., `"left"` → `"camera1"`).
+        defer_weight_load (bool): Build the exact policy `from_pretrained` would build — same
+            config resolution, same stats-derived buffers, same device placement and eval mode —
+            but skip the safetensors weight load. Used when resuming from a DCP checkpoint, whose
+            sharded weights stream in after `accelerator.prepare()` (the distributed checkpoint
+            engine overwrites the random init).

    Returns:
-        An instantiated and device-placed policy model.
+        PreTrainedPolicy: An instantiated and device-placed policy model.

    Raises:
        ValueError: If both or neither of `ds_meta` and `env_cfg` are provided.
-        NotImplementedError: If attempting to use an unsupported policy-backend
-                             combination (e.g., VQBeT with 'mps').
+        NotImplementedError: If attempting to use an unsupported policy-backend combination
+            (e.g., VQBeT with 'mps').
    """
    if bool(ds_meta) == bool(env_cfg):
        raise ValueError("Either one of a dataset metadata or a sim env must be provided.")
@@ -389,11 +338,18 @@ def make_policy(
        )

    if cfg.pretrained_path and not cfg.use_peft:
-        # Load a pretrained policy and override the config if needed (for example, if there are inference-time
-        # hyperparameters that we want to vary).
-        kwargs["pretrained_name_or_path"] = cfg.pretrained_path
-        kwargs["revision"] = cfg.pretrained_revision
-        policy = policy_cls.from_pretrained(**kwargs)
+        if defer_weight_load:
+            # Same construction path as from_pretrained (config already resolved from the
+            # checkpoint by the caller; dataset_stats/dataset_meta kwargs identical), minus the
+            # weight load — parity by construction.
+            policy = policy_cls(**kwargs)
+            policy.eval()
+        else:
+            # Load a pretrained policy and override the config if needed (for example, if there
+            # are inference-time hyperparameters that we want to vary).
+            kwargs["pretrained_name_or_path"] = cfg.pretrained_path
+            kwargs["revision"] = cfg.pretrained_revision
+            policy = policy_cls.from_pretrained(**kwargs)
    elif cfg.pretrained_path and cfg.use_peft:
        # Load a pretrained PEFT model on top of the policy. The pretrained path points to the folder/repo
        # of the adapter and the adapter's config contains the path to the base policy. So we need the
@@ -27,8 +27,6 @@ from lerobot.configs import (
 from lerobot.optim import AdamWConfig
 from lerobot.utils.constants import ACTION, OBS_STATE

-from ..rtc.configuration_rtc import RTCConfig
-
 WAN22_MODEL_ID = "Wan-AI/Wan2.2-TI2V-5B"
 WAN22_DIFFUSERS_MODEL_ID = "Wan-AI/Wan2.2-TI2V-5B-Diffusers"
 FASTWAM_BASE_MODEL_ID = "lerobot/fastwam_base"
@@ -190,25 +188,12 @@ class FastWAMConfig(PreTrainedConfig):
    action_video_freq_ratio: int = 4
    image_size: tuple[int, int] = (224, 448)
    context_len: int = 128
-
-    # Relative actions: converts absolute actions to relative (action -= state) during
-    # preprocessing, and reverses it at postprocessing. Requires `proprio_dim` (OBS_STATE).
-    use_relative_actions: bool = False
-    # Joint names to keep absolute (not converted to relative). Empty list = all dims relative.
-    relative_exclude_joints: list[str] = field(default_factory=lambda: ["gripper"])
-    # Populated at runtime from dataset metadata by make_policy (used to build the exclude mask).
-    action_feature_names: list[str] | None = None
    model_id: str = WAN22_MODEL_ID
    tokenizer_model_id: str = WAN_T5_TOKENIZER_ID
    text_encoder_model_id: str = WAN22_DIFFUSERS_MODEL_ID
    base_model_id: str | None = FASTWAM_BASE_MODEL_ID
    tokenizer_max_len: int = 128
    load_text_encoder: bool = True
-    # Device for the frozen ~11GB UMT5-XXL text encoder. `None` keeps it on the main
-    # policy `device` (default). Set to e.g. "cpu" to keep it off the GPU and save VRAM;
-    # prompts are then encoded on that device and the resulting embeddings moved to the
-    # policy device. Trades GPU memory for slower (CPU) text encoding.
-    text_encoder_device: str | None = None
    mot_checkpoint_mixed_attn: bool = False
    torch_dtype: str = "bfloat16"
    prompt_template: str = (
@@ -216,11 +201,6 @@ class FastWAMConfig(PreTrainedConfig):
    )
    num_inference_steps: int = 10
    inference_seed: int | None = 42
-    # Real-Time Chunking (RTC): async chunk generation with prefix guidance so a new chunk
-    # inpaints smoothly onto the still-executing tail of the previous one. `None` disables it
-    # (default synchronous inference). Consumed by `RTCInferenceEngine`, which calls
-    # `predict_action_chunk(..., inference_delay=, prev_chunk_left_over=)`.
-    rtc_config: RTCConfig | None = None
    rand_device: str = "cpu"
    text_cfg_scale: float = 1.0
    negative_prompt: str = ""
@@ -299,10 +279,6 @@ class FastWAMConfig(PreTrainedConfig):
        finally:
            self.pretrained_path = pretrained_path

-    @property
-    def chunk_size(self) -> int:
-        return self.action_horizon
-
    def get_optimizer_preset(self) -> AdamWConfig:
        return AdamWConfig(lr=self.optimizer_lr, weight_decay=self.optimizer_weight_decay)

@@ -22,7 +22,6 @@ import torch
 from torch import Tensor

 from lerobot.policies.pretrained import PreTrainedPolicy
-from lerobot.policies.rtc.modeling_rtc import RTCProcessor
 from lerobot.utils.constants import OBS_STATE
 from lerobot.utils.import_utils import require_package

@@ -55,6 +54,9 @@ class FastWAMPolicy(PreTrainedPolicy):

    config_class = FastWAMConfig
    name = "fastwam"
+    # FSDP2 wrap units: MoTLayer is the single FSDP owner of each layer's expert blocks
+    # (the blocks are re-parented onto it precisely so sharding has one boundary to hook).
+    _fsdp_wrap_modules = ["MoTLayer"]

    def __init__(
        self,
@@ -86,29 +88,8 @@ class FastWAMPolicy(PreTrainedPolicy):
                for layer in mot.layers:
                    if "video" in layer.blocks:
                        layer.blocks["video"].requires_grad_(False)
-        self.init_rtc_processor()
        self.reset()

-    def init_rtc_processor(self) -> None:
-        """Attach a Real-Time Chunking processor to the core model when configured.
-
-        Mirrors the PI0/PI05 pattern: the policy owns the `RTCProcessor` and hands it to
-        the core `FastWAM` model, which consults it inside `infer_action`'s denoising loop.
-        Must stay public and named exactly `init_rtc_processor`: the rollout loader
-        (`lerobot.rollout.context`) sets `policy.config.rtc_config = cfg.inference.rtc` and
-        then calls `policy.init_rtc_processor()` to (re)build the processor after load, so
-        `--inference.type=rtc` alone is enough to enable guidance — no separate policy-side
-        `rtc_config` needed. A private/renamed method would be silently skipped (guidance
-        off), degrading RTC to unguided async chunk-swapping.
-        """
-        self.rtc_processor = None
-        if self.config.rtc_config is not None:
-            self.rtc_processor = RTCProcessor(self.config.rtc_config)
-            self.model.rtc_processor = self.rtc_processor
-
-    def _rtc_enabled(self) -> bool:
-        return self.config.rtc_config is not None and self.config.rtc_config.enabled
-
    @classmethod
    def _load_as_safetensor(cls, model, model_file: str, map_location: str, strict: bool):
        """Shape-aware load that supports cross-embodiment fine-tuning.
@@ -172,24 +153,6 @@ class FastWAMPolicy(PreTrainedPolicy):

    def reset(self) -> None:
        self._action_queue: deque[Tensor] = deque([], maxlen=self.config.n_action_steps)
-        # Per-episode text-embedding cache (mirrors LingBot-VA's `_prompt_embeds`). The task
-        # is fixed for an episode, so the ~11GB UMT5 encoder runs once on the first chunk and
-        # the resulting context is reused for every subsequent chunk. Cleared here on reset so
-        # a new episode's (possibly different) task is re-encoded. Proprio is still appended
-        # fresh each chunk downstream, so only the text-only context is cached.
-        self._cached_prompt: Any = None
-        self._cached_context: Tensor | None = None
-        self._cached_context_mask: Tensor | None = None
-
-    def _encode_prompt_cached(self, prompt: Any) -> tuple[Tensor, Tensor]:
-        """Encode `prompt` to `(context, context_mask)`, reusing the cache when the prompt is
-        unchanged so UMT5 runs at most once per episode (per distinct task)."""
-        if self._cached_context is None or self._cached_prompt != prompt:
-            context, context_mask = self.model.encode_prompt(prompt)
-            self._cached_prompt = prompt
-            self._cached_context = context
-            self._cached_context_mask = context_mask
-        return self._cached_context, self._cached_context_mask

    def _batch_to_training_sample(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
        """Adapt a standard LeRobot batch to the FastWAM-native sample that
@@ -223,9 +186,7 @@ class FastWAMPolicy(PreTrainedPolicy):
                sample["proprio"] = state.unsqueeze(1) if state.ndim == 2 else state
        return sample

-    def forward(
-        self, batch: dict[str, Tensor], reduction: str = "mean"
-    ) -> tuple[Tensor, dict[str, Any]]:
+    def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict[str, Any]]:
        """Compute FastWAM training loss for a LeRobot batch.

        Args:
@@ -233,42 +194,24 @@ class FastWAMPolicy(PreTrainedPolicy):
                (`video`, `action`, `context`, `context_mask`) or LeRobot keys
                that can be adapted (`observation.images.*`, `observation.state`,
                `action`, `action_is_pad`).
-            reduction (str): "mean" returns the scalar loss (default, backward
-                compatible); "none" returns per-sample losses of shape (batch_size,)
-                for sample weighting (RA-BC).

        Returns:
-            tuple[Tensor, dict[str, Any]]: The loss to backprop (scalar for "mean",
-            per-sample (B,) for "none"), and a dict of logging metrics (e.g.
-            `loss_video`, `loss_action`) — the `(loss, output_dict)` contract the
-            LeRobot training loop expects.
+            tuple[Tensor, dict[str, Any]]: The scalar loss to backprop, and a dict of
+            logging metrics (e.g. `loss_video`, `loss_action`) — the `(loss, output_dict)`
+            contract the LeRobot training loop expects.
        """

        sample = self._batch_to_training_sample(batch)
-        loss, metrics = self.model.training_loss(sample, reduction=reduction)
+        loss, metrics = self.model.training_loss(sample)
        return loss, dict(metrics or {})

    @torch.no_grad()
-    def predict_action_chunk(
-        self,
-        batch: dict[str, Tensor],
-        inference_delay: int | None = None,
-        prev_chunk_left_over: Tensor | None = None,
-        execution_horizon: int | None = None,
-        **_: Any,
-    ) -> Tensor:
+    def predict_action_chunk(self, batch: dict[str, Tensor], **_: Any) -> Tensor:
        """Predict a chunk of actions from the current FastWAM observation.

        Args:
            batch (dict[str, Tensor]): Inference batch with `input_image` or
                image observation keys, plus `context/context_mask` or `prompt`.
-            inference_delay (int | None): RTC — number of prefix steps assumed already
-                executed by the time this chunk lands (from measured inference latency).
-            prev_chunk_left_over (Tensor | None): RTC — the previous chunk's unexecuted
-                action tail `[T_prev, action_dim]` in model space; guides denoising so the
-                new chunk inpaints onto it. `None` (default) = plain synchronous inference.
-            execution_horizon (int | None): RTC — override for the prefix-weight horizon;
-                `None` falls back to `rtc_config.execution_horizon`.

        Returns:
            Tensor: Action chunk with shape `[B, action_horizon, action_dim]`.
@@ -276,20 +219,6 @@ class FastWAMPolicy(PreTrainedPolicy):

        self.eval()
        infer_kwargs = _batch_to_infer_kwargs(batch=batch, config=self.config)
-        # Encode the task once per episode and reuse it (LingBot-VA parity): swap the raw
-        # `prompt` for the cached `context`/`context_mask` so `infer_action` skips `encode_prompt`
-        # and the text encoder isn't re-run every chunk. Skipped when the caller supplies its own
-        # precomputed `context` (the two are mutually exclusive downstream).
-        if infer_kwargs.get("context") is None and infer_kwargs.get("prompt") is not None:
-            context, context_mask = self._encode_prompt_cached(infer_kwargs["prompt"])
-            infer_kwargs["prompt"] = None
-            infer_kwargs["context"] = context
-            infer_kwargs["context_mask"] = context_mask
-        # RTC guidance args flow straight to `infer_action`; they are inert unless an
-        # RTCProcessor is attached, enabled, and `prev_chunk_left_over` is provided.
-        infer_kwargs["inference_delay"] = inference_delay
-        infer_kwargs["prev_chunk_left_over"] = prev_chunk_left_over
-        infer_kwargs["execution_horizon"] = execution_horizon
        batch_size = _infer_kwargs_batch_size(infer_kwargs)
        if batch_size == 1:
            action = _action_from_model_output(self.model.infer_action(**infer_kwargs))
@@ -334,10 +263,9 @@ class FastWAMPolicy(PreTrainedPolicy):
            mixtures={"video": video_expert, "action": action_expert},
            mot_checkpoint_mixed_attn=config.mot_checkpoint_mixed_attn,
        )
-        text_encoder_device = config.text_encoder_device or device
        text_encoder = (
            load_pretrained_wan_text_encoder(
-                model_id=config.text_encoder_model_id, torch_dtype=dtype, device=text_encoder_device
+                model_id=config.text_encoder_model_id, torch_dtype=dtype, device=device
            )
            if config.load_text_encoder
            else None
@@ -348,7 +276,6 @@ class FastWAMPolicy(PreTrainedPolicy):
            mot=mot,
            vae=load_pretrained_wan_vae(torch_dtype=dtype, device=device),
            text_encoder=text_encoder,
-            text_encoder_device=config.text_encoder_device,
            tokenizer=build_wan_tokenizer(
                model_id=config.tokenizer_model_id, tokenizer_max_len=config.tokenizer_max_len
            ),
@@ -21,13 +21,10 @@ import torch

 from lerobot.configs import PipelineFeatureType, PolicyFeature
 from lerobot.processor import (
-    AbsoluteActionsProcessorStep,
    ActionProcessorStep,
    PolicyAction,
    PolicyProcessorPipeline,
-    ProcessorStep,
    ProcessorStepRegistry,
-    RelativeActionsProcessorStep,
    make_default_policy_processor_steps,
    make_policy_processor_pipelines,
 )
@@ -101,25 +98,14 @@ def make_fastwam_pre_post_processors(

    steps = make_default_policy_processor_steps(config, normalization_stats, normalizer_device=config.device)

-    # Shared relative-action step (OpenPI order: raw -> relative -> normalize -> model ->
-    # unnormalize -> absolute). The SAME instance is passed to AbsoluteActionsProcessorStep
-    # below so its cached raw state (set during preprocessing) flows to postprocessing.
-    relative_step = RelativeActionsProcessorStep(
-        enabled=config.use_relative_actions,
-        exclude_joints=getattr(config, "relative_exclude_joints", []),
-        action_names=getattr(config, "action_feature_names", None),
-    )
-
-    input_steps: list[ProcessorStep] = [
+    input_steps = [
        steps.rename_observations,
        steps.add_batch_dim,
        steps.to_device,
-        relative_step,
        steps.normalize,
    ]
-    output_steps: list[ProcessorStep] = [
+    output_steps = [
        steps.unnormalize,
-        AbsoluteActionsProcessorStep(enabled=config.use_relative_actions, relative_step=relative_step),
    ]
    if config.toggle_action_dimensions:
        output_steps.append(
@@ -839,7 +839,6 @@ class FastWAM(torch.nn.Module):
        text_dim: int | None = None,
        proprio_dim: int | None = None,
        device: str = "cpu",
-        text_encoder_device: str | torch.device | None = None,
        torch_dtype: torch.dtype = torch.float32,
        video_train_shift: float = 5.0,
        video_infer_shift: float = 5.0,
@@ -908,27 +907,12 @@ class FastWAM(torch.nn.Module):
        self.train_scheduler = self.train_video_scheduler
        self.infer_scheduler = self.infer_video_scheduler

-        # Optional Real-Time Chunking processor (set by the policy wrapper). When present and
-        # enabled it guides the action denoising loop in `infer_action` so a freshly generated
-        # chunk inpaints onto the previous chunk's unexecuted tail. Plain attribute (not an
-        # nn.Module) — carries no parameters and stays out of state_dict / device moves.
-        self.rtc_processor = None
        self.device = torch.device(device)
-        # When pinned (e.g. "cpu"), the frozen text encoder stays on this device instead
-        # of following the model onto the GPU — `_apply` skips it and `encode_prompt` runs
-        # it here, moving embeddings back to `self.device`. `None` = follow `self.device`.
-        self._text_encoder_device = (
-            torch.device(text_encoder_device) if text_encoder_device is not None else None
-        )
        self.torch_dtype = torch_dtype
        self.loss_lambda_video = float(loss_lambda_video)
        self.loss_lambda_action = float(loss_lambda_action)

        self.to(self.device)
-        # `self.to` above (via `_apply`) skips a pinned text encoder; make sure it actually
-        # sits on the pinned device (it was loaded there, but this is a cheap safety net).
-        if self.text_encoder is not None and self._text_encoder_device is not None:
-            self.text_encoder._apply(lambda t: t.to(self._text_encoder_device))

    @classmethod
    def from_wan22_pretrained(
@@ -1019,8 +1003,7 @@ class FastWAM(torch.nn.Module):
        # while staying out of `state_dict()` / `parameters()`.
        super()._apply(fn, *args, **kwargs)
        self.vae._apply(fn)
-        # A pinned text encoder (e.g. on CPU) must NOT follow device moves — leave it put.
-        if self.text_encoder is not None and self._text_encoder_device is None:
+        if self.text_encoder is not None:
            self.text_encoder._apply(fn)
        return self

@@ -1041,12 +1024,9 @@ class FastWAM(torch.nn.Module):
                "Prompt encoding requires loaded text encoder/tokenizer. "
                "Set `load_text_encoder=true` or provide precomputed `context/context_mask`."
            )
-        # Run the encoder on its own device (may be pinned to CPU to save VRAM), then
-        # move the resulting embeddings/mask to the model device for the DiT.
-        te_device = self._text_encoder_device or self.device
        ids, mask = self.tokenizer(prompt, return_mask=True, add_special_tokens=True)
-        ids = ids.to(te_device)
-        mask = mask.to(te_device, dtype=torch.bool)
+        ids = ids.to(self.device)
+        mask = mask.to(self.device, dtype=torch.bool)
        prompt_emb = self.text_encoder(ids, mask)
        seq_lens = mask.gt(0).sum(dim=1).long()
        for i, v in enumerate(seq_lens):
@@ -1054,7 +1034,7 @@ class FastWAM(torch.nn.Module):
        # Match FastWAM/Wan2.2 context semantics: padding embeddings are zeroed,
        # while cross-attention still sees a fixed-length context.
        mask = torch.ones_like(mask)
-        return prompt_emb.to(device=self.device), mask.to(device=self.device)
+        return prompt_emb.to(device=self.device), mask

    def _append_proprio_to_context(
        self,
@@ -1379,9 +1359,7 @@ class FastWAM(torch.nn.Module):
        pred_action = self.action_expert.post_dit(tokens_out["action"], action_pre)
        return pred_video, pred_action

-    def _compute_training_video_loss(
-        self, inputs, pred_video, target_video, timestep_video, reduction: str = "mean"
-    ):
+    def _compute_training_video_loss(self, inputs, pred_video, target_video, timestep_video):
        include_initial_video_step = inputs["first_frame_latents"] is None
        if inputs["first_frame_latents"] is not None:
            pred_video = pred_video[:, :, 1:]
@@ -1396,13 +1374,9 @@ class FastWAM(torch.nn.Module):
            loss_video_per_sample.device,
            dtype=loss_video_per_sample.dtype,
        )
-        weighted = loss_video_per_sample * video_weight
-        # reduction="none" returns the per-sample vector (B,) for sample weighting (RA-BC).
-        return weighted if reduction == "none" else weighted.mean()
+        return (loss_video_per_sample * video_weight).mean()

-    def _compute_training_action_loss(
-        self, inputs, pred_action, target_action, timestep_action, reduction: str = "mean"
-    ):
+    def _compute_training_action_loss(self, inputs, pred_action, target_action, timestep_action):
        action_loss_token = functional.mse_loss(
            pred_action.float(), target_action.float(), reduction="none"
        ).mean(dim=2)
@@ -1419,11 +1393,9 @@ class FastWAM(torch.nn.Module):
            action_loss_per_sample.device,
            dtype=action_loss_per_sample.dtype,
        )
-        weighted = action_loss_per_sample * action_weight
-        # reduction="none" returns the per-sample vector (B,) for sample weighting (RA-BC).
-        return weighted if reduction == "none" else weighted.mean()
+        return (action_loss_per_sample * action_weight).mean()

-    def training_loss(self, sample, tiled: bool = False, reduction: str = "mean"):
+    def training_loss(self, sample, tiled: bool = False):
        inputs = self.build_inputs(sample, tiled=tiled)
        targets = self._sample_training_targets(inputs)
        pred_video, pred_action = self._run_training_mot(inputs=inputs, targets=targets)
@@ -1432,20 +1404,17 @@ class FastWAM(torch.nn.Module):
            pred_video=pred_video,
            target_video=targets["target_video"],
            timestep_video=targets["timestep_video"],
-            reduction=reduction,
        )
        loss_action = self._compute_training_action_loss(
            inputs=inputs,
            pred_action=pred_action,
            target_action=targets["target_action"],
            timestep_action=targets["timestep_action"],
-            reduction=reduction,
        )
-        # With reduction="none" both terms are (B,), so loss_total is the per-sample loss (B,).
        loss_total = self.loss_lambda_video * loss_video + self.loss_lambda_action * loss_action
        loss_dict = {
-            "loss_video": self.loss_lambda_video * float(loss_video.detach().mean().item()),
-            "loss_action": self.loss_lambda_action * float(loss_action.detach().mean().item()),
+            "loss_video": self.loss_lambda_video * float(loss_video.detach().item()),
+            "loss_action": self.loss_lambda_action * float(loss_action.detach().item()),
        }
        return loss_total, loss_dict

@@ -1830,9 +1799,6 @@ class FastWAM(torch.nn.Module):
        seed: int | None = None,
        rand_device: str = "cpu",
        tiled: bool = False,
-        inference_delay: int | None = None,
-        prev_chunk_left_over: torch.Tensor | None = None,
-        execution_horizon: int | None = None,
    ) -> dict[str, Any]:
        self.eval()
        if str(getattr(self.video_expert, "video_attention_mask_mode", "")) != "first_frame_causal":
@@ -1885,43 +1851,18 @@ class FastWAM(torch.nn.Module):
            dtype=latents_action.dtype,
            shift_override=sigma_shift,
        )
-        rtc_active = (
-            self.rtc_processor is not None
-            and getattr(self.rtc_processor.rtc_config, "enabled", False)
-            and prev_chunk_left_over is not None
-        )
-        num_train_timesteps = float(self.infer_action_scheduler.num_train_timesteps)
        for step_t_action, step_delta_action in zip(infer_timesteps_action, infer_deltas_action, strict=True):
            timestep_action = step_t_action.unsqueeze(0).to(dtype=latents_action.dtype, device=self.device)

-            def denoise(x_t, ts=timestep_action):
-                return self._predict_action_noise_with_cache(
-                    latents_action=x_t,
-                    timestep_action=ts,
-                    context=context,
-                    context_mask=context_mask,
-                    video_kv_cache=video_kv_cache,
-                    attention_mask=attention_mask,
-                    video_seq_len=video_seq_len,
-                )
-
-            if rtc_active:
-                # `time` is the flow-matching noise level sigma in [0, 1]: FastWAM's model
-                # predicts velocity v = noise - clean, so the clean-action estimate is
-                # x1 = x_t - sigma * v — exactly RTC's `x1_t = x_t - time * v_t`.
-                sigma = float(step_t_action.item()) / num_train_timesteps
-                pred_action = self.rtc_processor.denoise_step(
-                    x_t=latents_action,
-                    prev_chunk_left_over=prev_chunk_left_over.to(
-                        device=latents_action.device, dtype=latents_action.dtype
-                    ),
-                    inference_delay=inference_delay or 0,
-                    time=sigma,
-                    original_denoise_step_partial=denoise,
-                    execution_horizon=execution_horizon,
-                )
-            else:
-                pred_action = denoise(latents_action)
+            pred_action = self._predict_action_noise_with_cache(
+                latents_action=latents_action,
+                timestep_action=timestep_action,
+                context=context,
+                context_mask=context_mask,
+                video_kv_cache=video_kv_cache,
+                attention_mask=attention_mask,
+                video_seq_len=video_seq_len,
+            )

            latents_action = self.infer_action_scheduler.step(pred_action, step_delta_action, latents_action)

@@ -28,11 +28,7 @@ from dataclasses import dataclass, field
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
 from lerobot.optim.optimizers import AdamWConfig
-from lerobot.optim.schedulers import (
-    ConstantWithWarmupSchedulerConfig,
-    CosineAnnealingWithWarmupSchedulerConfig,
-    LRSchedulerConfig,
-)
+from lerobot.optim.schedulers import ConstantWithWarmupSchedulerConfig, LRSchedulerConfig
 from lerobot.utils.constants import ACTION


@@ -96,15 +92,6 @@ class LingBotVAConfig(PreTrainedConfig):
    # (un)normalization quantiles live in the checkpoint's ``policy_postprocessor.json``, not here.
    used_action_channel_ids: list[int] = field(default_factory=lambda: list(range(7)))

-    # Relative actions: converts absolute actions to relative (action -= state) during
-    # preprocessing, and reverses it at postprocessing. Requires the dataset to provide
-    # observation.state whose leading dims align 1:1 with the used action channels.
-    use_relative_actions: bool = False
-    # Joint names to keep absolute (not converted to relative). Empty list = all dims relative.
-    relative_exclude_joints: list[str] = field(default_factory=lambda: ["gripper"])
-    # Populated at runtime from dataset metadata by make_policy (used to build the exclude mask).
-    action_feature_names: list[str] | None = None
-
    # Opt-in: VAE-decode predicted video latents to ``self.last_predicted_frames`` for saving MP4s.
    save_predicted_video: bool = False

@@ -125,17 +112,6 @@ class LingBotVAConfig(PreTrainedConfig):
    optimizer_weight_decay: float = 1e-4
    optimizer_grad_clip_norm: float = 1.0
    scheduler_warmup_steps: int = 1000
-    # Scheduler after warmup. "constant_with_warmup" (upstream default: warmup then flat peak LR)
-    # or "cosine_annealing_with_warmup" (warmup then cosine anneal peak->0 over the remaining steps).
-    # Cosine tightens the loss tail and often nudges final loss down; it does NOT reduce the
-    # flow-matching estimator's step-to-step noise (that's metric variance, LR-independent).
-    scheduler_type: str = "constant_with_warmup"
-    # Probability of corrupting the action stream's conditioning (clean/context) tokens with
-    # flow-matching noise during training, mirroring the video stream's noisy_cond_prob=0.5.
-    # Upstream train.py hardcodes 0.0 for actions (never corrupted) with no exposed knob; this is
-    # an experimental deviation to make the model more tolerant of imperfect action history
-    # (e.g. clamp-induced drift between predicted and executed actions during rollout).
-    action_noisy_cond_prob: float = 0.0

    def __post_init__(self):
        super().__post_init__()
@@ -174,32 +150,14 @@ class LingBotVAConfig(PreTrainedConfig):
        )

    def get_scheduler_preset(self) -> LRSchedulerConfig | None:
-        # Default (upstream): linear warmup then constant LR (warmup_constant_lambda).
-        # Optionally cosine-anneal peak->0 over the remaining steps via scheduler_type.
-        if self.scheduler_type == "cosine_annealing_with_warmup":
-            return CosineAnnealingWithWarmupSchedulerConfig(num_warmup_steps=self.scheduler_warmup_steps)
+        # Upstream uses a linear warmup followed by a constant LR (warmup_constant_lambda).
        return ConstantWithWarmupSchedulerConfig(num_warmup_steps=self.scheduler_warmup_steps)

    @property
    def observation_delta_indices(self) -> list[int]:
-        """Observation frame deltas for the training clip, sized to what the VAE actually reads.
-
-        ``diffusers``' ``AutoencoderKLWan._encode`` runs ``iter_ = 1 + (n - 1) // 4`` passes over
-        ``x[:, :, :1]`` then ``x[:, :, 1 + 4*(i-1) : 1 + 4*i]``, so it only ever consumes the first
-        ``4 * (iter_ - 1) + 1`` frames of an ``n``-frame clip. Asking for ``frame_chunk_size * 4``
-        frames (the previous formula) therefore decoded 3 frames per sample that never reached the
-        encoder: at ``frame_chunk_size=2`` the deltas were ``[0, 4, ..., 28]`` and only
-        ``[0, 4, 8, 12, 16]`` were used -- verified by ablation, scrambling the tail left the latents
-        bit-identical.
-
-        Requesting exactly ``4 * (frame_chunk_size - 1) + 1`` frames yields the same
-        ``frame_chunk_size`` latent frames with every loaded frame used, and drops the wasted video
-        decode. The stride is unchanged, so the frames that do reach the model are the same ones.
-        """
        temporal_downsample = 4
        stride = max(1, self.action_per_frame // temporal_downsample)
-        num_frames = temporal_downsample * (self.frame_chunk_size - 1) + 1
-        return [i * stride for i in range(num_frames)]
+        return list(range(0, self.frame_chunk_size * temporal_downsample * stride, stride))

    @property
    def action_delta_indices(self) -> list[int]:
@@ -38,7 +38,7 @@ import torch.nn.functional as F  # noqa: N812
 from einops import rearrange
 from torch import Tensor

-from lerobot.policies.pretrained import PreTrainedPolicy, unpack_action_output
+from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.utils.constants import ACTION
 from lerobot.utils.import_utils import require_package

@@ -99,6 +99,8 @@ class LingBotVAPolicy(PreTrainedPolicy):
        # from ``config.wan_pretrained_path`` the first time inference runs.
        self._frozen: dict = {}

+        self.last_predicted_frames: Tensor | None = None
+        self.last_predicted_latents: Tensor | None = None
        self.reset()

    # Frozen-module lazy loading (VAE + UMT5 + tokenizer)
@@ -168,6 +170,8 @@ class LingBotVAPolicy(PreTrainedPolicy):
        self._prompt: str | None = None
        self._prompt_embeds = None
        self._negative_prompt_embeds = None
+        self.last_predicted_frames = None
+        self.last_predicted_latents = None
        self._use_cfg = (cfg.guidance_scale > 1) or (cfg.action_guidance_scale > 1)
        # Two independent flow-matching schedulers (video latent + action streams).
        self._scheduler = FlowMatchScheduler(shift=cfg.snr_shift, sigma_min=0.0, extra_one_step=True)
@@ -253,12 +257,8 @@ class LingBotVAPolicy(PreTrainedPolicy):
            "grid_id": grid_id,
        }

-    def _flow_matching_loss(self, input_dict, pred, reduction: str = "mean"):
-        """Dual-stream flow-matching loss (port of upstream ``Trainer.compute_loss``).
-
-        ``reduction="mean"`` returns scalar (latent_loss, action_loss); ``"none"`` returns
-        per-sample vectors of shape ``(B,)`` each (averaged over latent frames) for RA-BC.
-        """
+    def _flow_matching_loss(self, input_dict, pred):
+        """Dual-stream flow-matching loss (port of upstream ``Trainer.compute_loss``)."""
        latent_pred, action_pred = pred
        ld, ad = input_dict["latent_dict"], input_dict["action_dict"]
        action_pred = rearrange(action_pred, "b (f n) c -> b c f n 1", f=ad["targets"].shape[-3])
@@ -278,8 +278,7 @@ class LingBotVAPolicy(PreTrainedPolicy):
        latent_loss = (
            (latent_loss * lw[:, None, :, None, None]).permute(0, 2, 3, 4, 1).flatten(0, 1).flatten(1)
        )
-        # per (batch*frame) mean over spatial/channel -> (B*F,)
-        latent_loss = latent_loss.sum(dim=1) / (torch.ones_like(latent_loss).sum(dim=1) + 1e-6)
+        latent_loss = (latent_loss.sum(dim=1) / (torch.ones_like(latent_loss).sum(dim=1) + 1e-6)).mean()

        amask = ad["actions_mask"].float()
        action_loss = F.mse_loss(action_pred.float(), ad["targets"].float().detach(), reduction="none")
@@ -287,14 +286,10 @@ class LingBotVAPolicy(PreTrainedPolicy):
            (action_loss * aw[:, None, :, None, None] * amask).permute(0, 2, 3, 4, 1).flatten(0, 1).flatten(1)
        )
        amask_f = amask.permute(0, 2, 3, 4, 1).flatten(0, 1).flatten(1)
-        action_loss = action_loss.sum(dim=1) / (amask_f.sum(dim=1) + 1e-6)
+        action_loss = (action_loss.sum(dim=1) / (amask_f.sum(dim=1) + 1e-6)).mean()
+        return latent_loss, action_loss

-        if reduction == "none":
-            # (B*F,) -> (B, F) -> (B,): per-sample losses for RA-BC weighting.
-            return latent_loss.reshape(bn, fn).mean(dim=1), action_loss.reshape(bn, fn).mean(dim=1)
-        return latent_loss.mean(), action_loss.mean()
-
-    def training_loss_from_streams(self, latents, actions, actions_mask, text_emb, reduction: str = "mean"):
+    def training_loss_from_streams(self, latents, actions, actions_mask, text_emb):
        """Core dual-stream training loss given prepared latents / actions / text embeddings.

        ``latents``: ``[B, in_channels, F, h, w]`` (normalized video latents).
@@ -311,11 +306,7 @@ class LingBotVAPolicy(PreTrainedPolicy):
            latents, self._train_sched_latent, action_mask=None, action_mode=False, noisy_cond_prob=0.5
        )
        action_dict = self._add_noise_stream(
-            actions,
-            self._train_sched_action,
-            action_mask=actions_mask,
-            action_mode=True,
-            noisy_cond_prob=self.config.action_noisy_cond_prob,
+            actions, self._train_sched_action, action_mask=actions_mask, action_mode=True, noisy_cond_prob=0.0
        )
        latent_dict["text_emb"] = text_emb
        action_dict["text_emb"] = text_emb
@@ -327,24 +318,20 @@ class LingBotVAPolicy(PreTrainedPolicy):
            "window_size": int(torch.randint(4, 65, (1,)).item()),
        }
        pred = self.transformer(input_dict, train_mode=True)
-        latent_loss, action_loss = self._flow_matching_loss(input_dict, pred, reduction)
-        # reduction="none": latent_loss/action_loss are (B,) -> loss is per-sample (B,).
+        latent_loss, action_loss = self._flow_matching_loss(input_dict, pred)
        loss = latent_loss + action_loss
-        return loss, {"latent_loss": latent_loss.mean().item(), "action_loss": action_loss.mean().item()}
+        return loss, {"latent_loss": latent_loss.item(), "action_loss": action_loss.item()}

-    def forward(self, batch: dict[str, Tensor], reduction: str = "mean") -> tuple[Tensor, dict | None]:
+    def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict | None]:
        """Training forward: dual-stream flow-matching loss.

        Builds the (video-latent, action, text) training streams from a LeRobot batch
        (VAE-encoding the camera frames and UMT5-encoding the task), then runs the flow-matching
        dual-stream loss. Requires the policy to be built with ``attn_mode='flex'``.
-
-        ``reduction="mean"`` returns the scalar loss (default); ``"none"`` returns per-sample
-        losses of shape ``(B,)`` for sample weighting (RA-BC).
        """
        self._ensure_frozen_modules()
        latents, actions, actions_mask, text_emb = self._build_training_streams(batch)
-        return self.training_loss_from_streams(latents, actions, actions_mask, text_emb, reduction=reduction)
+        return self.training_loss_from_streams(latents, actions, actions_mask, text_emb)

    @torch.no_grad()
    def _build_training_streams(self, batch):
@@ -413,31 +400,22 @@ class LingBotVAPolicy(PreTrainedPolicy):
        return torch.cat(per_cam, dim=-1).to(self.config.device)

    @torch.no_grad()
-    def select_action(
-        self, batch: dict[str, Tensor], return_intermediate_predictions: bool = False, **kwargs
-    ) -> Tensor | tuple[Tensor, dict[str, Tensor]]:
+    def select_action(self, batch: dict[str, Tensor], **kwargs) -> Tensor:
        """Return one action, refilling the chunk (and feeding back observed keyframes) as needed.

        Mirrors the upstream LIBERO client loop (``evaluation/libero/client.py``): the first obs is
        the conditioning frame; every observation produced afterwards is buffered as a keyframe and,
        once the chunk's actions are exhausted, the buffered frames + executed actions are fed back
        into the KV cache before the next chunk is predicted.
-
-        When ``return_intermediate_predictions=True`` returns ``(action, predictions)``. Predictions
-        are produced only on the ticks that predict a fresh chunk (first tick and each chunk refill);
-        on the intermediate ticks that just pop a cached action, ``predictions`` is an empty dict.
        """
        self.eval()
        self._ensure_frozen_modules()
        self._maybe_init_prompt(batch)

-        predictions: dict[str, Tensor] = {}
        if not self._started:
            # First call: this observation conditions the first chunk (it is *not* a keyframe).
            self._started = True
-            actions, predictions = unpack_action_output(
-                self.predict_action_chunk(batch, return_intermediate_predictions=return_intermediate_predictions)
-            )  # [B, chunk_size, n_used]
+            actions = self.predict_action_chunk(batch)  # [B, chunk_size, n_used]
            self._action_queue.extend(actions.transpose(0, 1))  # [chunk_size, B, n_used]
            self._obs_buffer = []
            self._exec_step = 0
@@ -449,31 +427,17 @@ class LingBotVAPolicy(PreTrainedPolicy):
            if len(self._action_queue) == 0:
                # All actions for the current chunk have been executed; feed the observed
                # keyframes + executed actions back and predict the next chunk.
-                actions, predictions = unpack_action_output(
-                    self.predict_action_chunk(
-                        None, return_intermediate_predictions=return_intermediate_predictions
-                    )
-                )
+                actions = self.predict_action_chunk(None)
                self._action_queue.extend(actions.transpose(0, 1))
                self._exec_step = 0

        self._prev_j = self._exec_step % self.config.action_per_frame
        self._exec_step += 1
-        action = self._action_queue.popleft()
-        if return_intermediate_predictions:
-            return action, predictions
-        return action
+        return self._action_queue.popleft()

    @torch.no_grad()
-    def predict_action_chunk(
-        self, batch: dict[str, Tensor], return_intermediate_predictions: bool = False, **kwargs
-    ) -> Tensor | tuple[Tensor, dict[str, Tensor]]:
-        """Run one autoregressive chunk and return actions ``[B, chunk_size, n_used]`` (normalized).
-
-        When ``return_intermediate_predictions=True`` returns ``(actions, predictions)`` where
-        ``predictions`` holds this chunk's VAE-decoded imagined video under ``"images.predicted"``
-        (``[T, H, W, 3]`` uint8 on CPU).
-        """
+    def predict_action_chunk(self, batch: dict[str, Tensor], **kwargs) -> Tensor:
+        """Run one autoregressive chunk and return actions ``[B, chunk_size, n_used]`` (normalized)."""
        self.eval()
        self._ensure_frozen_modules()
        self._maybe_init_prompt(batch)
@@ -495,6 +459,12 @@ class LingBotVAPolicy(PreTrainedPolicy):
        # actions: [B, action_dim, F, action_per_frame, 1] (model-normalized). Keep for KV feedback.
        self._executed_actions = actions

+        if self.config.save_predicted_video:
+            # Match upstream LingBot-VA visualization: collect chunk latents and decode the
+            # concatenated latent sequence once after the rollout finishes.
+            self.last_predicted_frames = None
+            self.last_predicted_latents = latents.detach().to("cpu")
+
        # On the first chunk, frame 0 is the conditioning frame (already "known"): the upstream
        # LIBERO client skips it (start_idx=1), so we drop the first frame's actions here.
        used = self.config.used_action_channel_ids
@@ -503,15 +473,7 @@ class LingBotVAPolicy(PreTrainedPolicy):
            a = a[:, :, 1:]  # drop frame 0 -> (F-1) frames of actions
        a = a.squeeze(-1).flatten(2)  # [B, n_used, n_steps]
        a = a.transpose(1, 2).contiguous()  # [B, n_steps, n_used]
-        a = a.to(torch.float32)
-
-        if return_intermediate_predictions:
-            # Decode this chunk's imagined video for visualization / eval. Per-chunk decode (the VAE
-            # has no streaming decoder) may differ slightly at chunk boundaries from a single decode
-            # over the whole concatenated latent sequence; acceptable for monitoring/inspection.
-            frames = self._decode_predicted_video(latents)  # [T, H, W, 3] uint8, CPU
-            return a, {"images.predicted": frames}
-        return a
+        return a.to(torch.float32)

    # Prompt / text encoding
    def _maybe_init_prompt(self, batch):
@@ -872,6 +834,11 @@ class LingBotVAPolicy(PreTrainedPolicy):
        return actions, latents

    # Predicted-video decoding (opt-in)
+    @torch.no_grad()
+    def decode_predicted_latents(self, latents) -> Tensor:
+        """Decode a concatenated predicted-latent sequence into ``[T, H, W, 3]`` uint8 frames."""
+        return self._decode_predicted_video(latents)
+
    @torch.no_grad()
    def _decode_predicted_video(self, latents) -> Tensor:
        """VAE-decode predicted latents into a uint8 frame stack ``[T, H, W, 3]`` on CPU."""
@@ -25,11 +25,9 @@ import torch

 from lerobot.configs.types import FeatureType, NormalizationMode
 from lerobot.processor import (
-    AbsoluteActionsProcessorStep,
    PolicyAction,
    PolicyProcessorPipeline,
    ProcessorStep,
-    RelativeActionsProcessorStep,
    UnnormalizerProcessorStep,
    make_default_policy_processor_steps,
    make_policy_processor_pipelines,
@@ -49,33 +47,20 @@ def make_lingbot_va_pre_post_processors(

    steps = make_default_policy_processor_steps(config, dataset_stats)

-    # Shared relative-action step (OpenPI order: raw -> relative -> normalize -> model ->
-    # unnormalize -> absolute). The SAME instance is passed to AbsoluteActionsProcessorStep
-    # below so its cached raw state (set during preprocessing) flows to postprocessing.
-    relative_step = RelativeActionsProcessorStep(
-        enabled=config.use_relative_actions,
-        exclude_joints=getattr(config, "relative_exclude_joints", []),
-        action_names=getattr(config, "action_feature_names", None),
-    )
-
    input_steps: list[ProcessorStep] = [
        steps.rename_observations,
        steps.add_batch_dim,
-        relative_step,
        steps.normalize,
        steps.to_device,
    ]

-    # Unnormalize actions back to physical units. Config-driven norm_map (was hardcoded QUANTILES)
-    # so it stays symmetric with the preprocessor's NormalizerProcessorStep — required for
-    # use_relative_actions with ACTION=IDENTITY (and unchanged for QUANTILES runs).
+    # Unnormalize actions from [-1, 1] to physical units (QUANTILES) using q01/q99 restored from the checkpoint.
    output_steps: list[ProcessorStep] = [
        UnnormalizerProcessorStep(
            features=config.output_features,
-            norm_map=config.normalization_mapping,
+            norm_map={FeatureType.ACTION: NormalizationMode.QUANTILES},
            stats=dataset_stats,
        ),
-        AbsoluteActionsProcessorStep(enabled=config.use_relative_actions, relative_step=relative_step),
        steps.to_cpu,
    ]

@@ -604,6 +604,12 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
    ) -> torch.Tensor:
        """
        Optimized autoregressive decoding for FAST tokens using KV Caching.
+
+        Greedy decoding stops once every sequence emits the end-of-action marker. The
+        returned tensor keeps its fixed shape, with positions not generated after the
+        batch-wide stop left zero-filled. Stochastic decoding always runs to
+        ``max_decoding_steps`` so early stopping does not change the RNG state used by
+        subsequent calls.
        """
        if max_decoding_steps is None:
            max_decoding_steps = self.config.max_action_tokens
@@ -612,6 +618,12 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
        device = tokens.device
        lm_head = self.paligemma_with_expert.paligemma.lm_head

+        # detokenize_actions() cuts at the first "|", so greedy decoding can stop once
+        # every sequence has emitted it. Keep stochastic decoding unchanged because
+        # skipping multinomial calls would shift the RNG state for subsequent calls.
+        end_of_action_token_id = self._paligemma_tokenizer.convert_tokens_to_ids("|")
+        finished = torch.zeros(bsize, dtype=torch.bool, device=device) if temperature == 0 else None
+
        # --- 1. PREFILL PHASE ---
        # Process Images + Text Prompt + BOS token once to populate the KV cache.

@@ -663,6 +675,10 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
        # Initialize storage for generated tokens
        generated_action_tokens = torch.zeros((bsize, max_decoding_steps), dtype=torch.long, device=device)
        generated_action_tokens[:, 0] = next_token.squeeze(-1)
+        if finished is not None:
+            finished |= next_token.squeeze(-1) == end_of_action_token_id
+            if bool(finished.all()):
+                return generated_action_tokens

        # Track valid tokens mask (0 for pad, 1 for valid)
        # We need this to tell the new token what it can attend to (images + text + past actions)
@@ -713,6 +729,11 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`

            generated_action_tokens[:, t] = next_token.squeeze(-1)

+            if finished is not None:
+                finished |= next_token.squeeze(-1) == end_of_action_token_id
+                if bool(finished.all()):
+                    break
+
        return generated_action_tokens


@@ -18,20 +18,18 @@ import builtins
 import dataclasses
 import logging
 import os
-from importlib.resources import files
+import warnings
 from pathlib import Path
-from tempfile import TemporaryDirectory
-from typing import TYPE_CHECKING, TypedDict, TypeVar, Unpack
+from typing import TYPE_CHECKING, Any, ClassVar, TypedDict, TypeVar, Unpack

-from huggingface_hub import HfApi, ModelCard, ModelCardData, hf_hub_download, save_torch_state_dict
+from huggingface_hub import hf_hub_download, save_torch_state_dict
 from huggingface_hub.constants import SAFETENSORS_SINGLE_FILE
 from huggingface_hub.errors import HfHubHTTPError
-from safetensors.torch import load_model as load_model_as_safetensor, save_model as save_model_as_safetensor
+from safetensors.torch import load_model as load_model_as_safetensor
 from torch import Tensor, nn

-from lerobot.__version__ import __version__
 from lerobot.configs import PreTrainedConfig
-from lerobot.configs.train import TrainPipelineConfig
+from lerobot.utils.constants import ACTION
 from lerobot.utils.device_utils import resolve_safetensors_device
 from lerobot.utils.hub import HubMixin
 from lerobot.utils.import_utils import _peft_available, require_package
@@ -46,72 +44,18 @@ else:
    get_peft_model = None

 if TYPE_CHECKING:
+    from lerobot.configs.train import TrainPipelineConfig
    from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata

 T = TypeVar("T", bound="PreTrainedPolicy")

-
-def _build_card_context(
-    cfg: TrainPipelineConfig | None,
-    dataset_meta: LeRobotDatasetMetadata | None,
-    input_features: dict | None,
-    output_features: dict | None,
-) -> dict:
-    """Collect optional data for the model-card template.
-
-    Returns plain values only (no Markdown) — the template in
-    ``lerobot/templates/lerobot_modelcard_template.md`` decides how and whether to show
-    each one. Everything is best-effort: anything unavailable is left empty/None and the
-    template simply skips that section, so this never breaks a Hub push.
-    """
-    context = {
-        "training": None,
-        "input_features": input_features or {},
-        "output_features": output_features or {},
-        "dataset": None,
-        "robot_type": None,
-        "cameras": [],
-    }
-
-    if cfg is not None:
-        optimizer = getattr(cfg, "optimizer", None)
-        context["training"] = {
-            "steps": cfg.steps,
-            "batch_size": cfg.batch_size,
-            "seed": cfg.seed,
-            "optimizer": getattr(optimizer, "type", None) if optimizer else None,
-            "lr": getattr(optimizer, "lr", None) if optimizer else None,
-            "lerobot_version": __version__,
-        }
-
-    if dataset_meta is not None:
-        context["dataset"] = {
-            "repo_id": dataset_meta.repo_id,
-            "episodes": dataset_meta.total_episodes,
-            "frames": dataset_meta.total_frames,
-            "fps": dataset_meta.fps,
-            "tasks": [str(task) for task in dataset_meta.tasks.index],
-        }
-        context["robot_type"] = dataset_meta.robot_type
-        context["cameras"] = [key.split(".")[-1] for key in dataset_meta.camera_keys]
-
-    return context
+# Pinned far above any policy's total size so save_torch_state_dict always emits exactly one
+# `model.safetensors` (no shards, no index) — a constant, not a computed byte count.
+_SINGLE_FILE_SHARD_SIZE = "1TB"


 class ActionSelectKwargs(TypedDict, total=False):
    noise: Tensor | None
-    return_intermediate_predictions: bool
-
-
-def unpack_action_output(out: Tensor | tuple[Tensor, dict[str, Tensor]]) -> tuple[Tensor, dict[str, Tensor]]:
-    """Normalize a ``select_action`` / ``predict_action_chunk`` return to ``(action, predictions)``.
-
-    These methods return a bare action ``Tensor`` by default, or a ``(action, predictions)`` tuple when
-    called with ``return_intermediate_predictions=True``. A bare tensor becomes ``(tensor, {})``.
-    """
-    if isinstance(out, tuple):
-        return out[0], out[1]
-    return out, {}


 class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
@@ -122,6 +66,22 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
    config_class: None
    name: None

+    # --- declarative parallelism/acceleration surface ----------------------------------------
+    # Module CLASS names forming the FSDP2 wrap units (and, once wired, the activation-
+    # checkpointing units). Resolved onto the accelerate plugin right before
+    # `accelerator.prepare()` by `lerobot.distributed.set_fsdp_wrap_modules`; sharded training
+    # with no wrap source anywhere fails loudly instead of silently wrapping only the root.
+    _fsdp_wrap_modules: ClassVar[list[str] | None] = None
+    # Non-`forward` entry points that must trigger FSDP2 unshard/reshard hooks when called on a
+    # sharded policy (registered post-prepare via `torch.distributed.fsdp
+    # .register_fsdp_forward_method`); calling them unregistered crashes on mixed Tensor/DTensor.
+    _fsdp_forward_methods: ClassVar[tuple[str, ...]] = ("select_action", "predict_action_chunk")
+    # Capability gate for the (future) activation-checkpointing wiring.
+    supports_gradient_checkpointing: ClassVar[bool] = False
+    # Declarative context-parallel plan (diffusers `ContextParallelModelPlan` semantics:
+    # module FQN -> sequence split/gather spec). Reserved for the CP engine round.
+    _cp_plan: ClassVar[dict[str, Any] | None] = None
+
    def __init__(self, config: PreTrainedConfig, *inputs, **kwargs):
        super().__init__()
        if not isinstance(config, PreTrainedConfig):
@@ -139,43 +99,33 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
        if not getattr(cls, "name", None):
            raise TypeError(f"Class {cls.__name__} must define 'name'")

-    def save_pretrained(
-        self,
-        save_directory: str | Path,
-        *,
-        state_dict: dict[str, Tensor] | None = None,
-        repo_id: str | None = None,
-        push_to_hub: bool = False,
-        card_kwargs: dict | None = None,
-        **push_to_hub_kwargs,
-    ) -> str | None:
-        """Save the policy to a directory (and optionally push to the Hub).
+    def _save_pretrained(self, save_directory: Path) -> None:
+        """Serialize this policy's parameters (and config) into `save_directory`.

-        Overrides `HubMixin.save_pretrained` to add a `state_dict` argument (mirroring
-        `transformers.PreTrainedModel.save_pretrained`). Under FSDP, `self.state_dict()` would
-        return sharded tensors, so the caller gathers the full state dict via a cross-rank
-        collective and passes it here for `_save_pretrained` to write directly.
+        Sharding is handled internally: under FSDP2 the full state dict is gathered through a
+        COLLECTIVE, so when the policy is sharded this method (via `save_pretrained`) must be
+        called on EVERY rank — a rank-0-gated call deadlocks. File writes happen on the main
+        process only, in all layouts (single, DDP, sharded).
+
+        Args:
+            save_directory (Path): Target directory for the policy config (`config.json`) and the
+                safetensors weight file(s).
        """
-        save_directory = Path(save_directory)
-        save_directory.mkdir(parents=True, exist_ok=True)
-        self._save_pretrained(save_directory, state_dict=state_dict)
-        if push_to_hub:
-            if repo_id is None:
-                repo_id = save_directory.name
-            return self.push_to_hub(repo_id=repo_id, card_kwargs=card_kwargs, **push_to_hub_kwargs)
-        return None
+        # Lazy imports: the persistence layer pulls in lerobot.distributed only when saving.
+        from lerobot.distributed.checkpoint import full_model_state_dict, is_sharded_module
+        from lerobot.distributed.utils import is_main_process

-    def _save_pretrained(self, save_directory: Path, state_dict: dict[str, Tensor] | None = None) -> None:
-        self.config._save_pretrained(save_directory)
        model_to_save = self.module if hasattr(self, "module") else self
-        if state_dict is None:
-            save_model_as_safetensor(model_to_save, str(save_directory / SAFETENSORS_SINGLE_FILE))
+        if is_sharded_module(model_to_save):
+            logging.info("Gathering the full state dict from all ranks (sharded policy).")
+        state_dict = full_model_state_dict(model_to_save)  # collective when sharded; {} off-main
+        if not state_dict or not is_main_process():
+            # Sharded: the gather materializes on the main rank only (emptiness check).
+            # Non-sharded multi-rank (DDP): every rank holds a full dict — the explicit rank
+            # gate prevents N ranks racing on the same files. Single process: never taken.
            return
-        # A pre-gathered (e.g. FSDP full) state dict was supplied: write it directly.
-        # `save_torch_state_dict` discards shared-tensor duplicates just like `save_model` does;
-        # pin `max_shard_size` above the total size so the output stays a single `model.safetensors`
-        total_bytes = sum(t.numel() * t.element_size() for t in state_dict.values())
-        save_torch_state_dict(state_dict, str(save_directory), max_shard_size=max(total_bytes, 1))
+        self.config._save_pretrained(save_directory)
+        save_torch_state_dict(state_dict, str(save_directory), max_shard_size=_SINGLE_FILE_SHARD_SIZE)

    @classmethod
    def from_pretrained(
@@ -261,6 +211,29 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
        """
        raise NotImplementedError

+    def drop_queued_actions(self) -> None:
+        """Discard actions precomputed by earlier ``select_action`` calls.
+
+        Chunking policies answer most control ticks from a queue filled by an
+        earlier forward pass, so a mid-episode change to the conditioning —
+        e.g. a new language instruction — would otherwise only take effect
+        once that queue drains (up to ``chunk_size`` ticks).  Dropping the
+        queue forces a fresh forward pass on the next ``select_action``.
+
+        Unlike :meth:`reset` this keeps the rest of the episode state (e.g.
+        observation history), so it does not perturb policies that condition
+        on it.  Call it from the thread that calls ``select_action``: it
+        mutates the same queues that thread pops from.
+
+        Policies that keep no action queue inherit a no-op.
+        """
+        queues = getattr(self, "_queues", None)
+        if isinstance(queues, dict) and ACTION in queues:
+            queues[ACTION].clear()
+        action_queue = getattr(self, "_action_queue", None)
+        if action_queue is not None:
+            action_queue.clear()
+
    def supports_rtc(self) -> bool:
        """Whether this policy implements Real-Time Chunking inference semantics."""
        return False
@@ -280,34 +253,20 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
        raise NotImplementedError

    @abc.abstractmethod
-    def predict_action_chunk(
-        self, batch: dict[str, Tensor], **kwargs: Unpack[ActionSelectKwargs]
-    ) -> Tensor | tuple[Tensor, dict[str, Tensor]]:
+    def predict_action_chunk(self, batch: dict[str, Tensor], **kwargs: Unpack[ActionSelectKwargs]) -> Tensor:
        """Returns the action chunk (for action chunking policies) for a given observation, potentially in batch mode.

        Child classes using action chunking should use this method within `select_action` to form the action chunk
        cached for selection.
-
-        By default returns just the action `Tensor`. If `return_intermediate_predictions=True`,
-        returns `(action, predictions)` where `predictions` is a (possibly empty) `dict[str, Tensor]`
-        of additional model predictions a policy may expose (e.g. world-model predicted frames).
-        Policies that produce nothing extra may ignore the kwarg.
        """
        raise NotImplementedError

    @abc.abstractmethod
-    def select_action(
-        self, batch: dict[str, Tensor], **kwargs: Unpack[ActionSelectKwargs]
-    ) -> Tensor | tuple[Tensor, dict[str, Tensor]]:
+    def select_action(self, batch: dict[str, Tensor], **kwargs: Unpack[ActionSelectKwargs]) -> Tensor:
        """Return one action to run in the environment (potentially in batch mode).

        When the model uses a history of observations, or outputs a sequence of actions, this method deals
        with caching.
-
-        By default returns just the action `Tensor`. If `return_intermediate_predictions=True`,
-        returns `(action, predictions)` where `predictions` is a (possibly empty) `dict[str, Tensor]`
-        of additional model predictions a policy may expose (e.g. world-model predicted frames).
-        Policies that produce nothing extra may ignore the kwarg.
        """
        raise NotImplementedError

@@ -317,92 +276,39 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
        peft_model=None,
        state_dict: dict[str, Tensor] | None = None,
        dataset_meta: LeRobotDatasetMetadata | None = None,
-    ):
-        api = HfApi()
-        repo_id = api.create_repo(
-            repo_id=self.config.repo_id, private=self.config.private, exist_ok=True
-        ).repo_id
+    ) -> None:
+        """Publish this policy to the Hub.

-        # Push the files to the repo in a single commit
-        with TemporaryDirectory(ignore_cleanup_errors=True) as tmp:
-            saved_path = Path(tmp) / repo_id
+        Deprecated: use :func:`lerobot.common.train_utils.publish_trained_model` instead, which
+        also publishes the pre/post-processors alongside the model.

-            if peft_model is not None:
-                # Since PEFT just forwards calls to `push_model_to_hub`, `self` is not the PeftModel wrapper
-                # but the actual policy which is why we need the PEFT model passed to us to save the adapter.
-                # That also means that we need to store the policy config ourselves since PEFT can't.
-                peft_model.save_pretrained(saved_path)
-                self.config.save_pretrained(saved_path)
-            else:
-                # Calls _save_pretrained and stores model tensors
-                self.save_pretrained(saved_path, state_dict=state_dict)
+        Args:
+            cfg (TrainPipelineConfig): The training config; saved as `train_config.json` and
+                used to render the model card.
+            peft_model: The PEFT wrapper when training adapters, whose weights replace the full
+                model weights in the published repo. Defaults to None.
+            state_dict (dict[str, Tensor] | None): Ignored; weights are now gathered internally
+                when the policy is sharded. Defaults to None.
+            dataset_meta (LeRobotDatasetMetadata | None): Dataset metadata for the model card,
+                if available. Defaults to None.
+        """
+        from lerobot.common.train_utils import publish_trained_model

-            card = self.generate_model_card(
-                cfg.dataset.repo_id,
-                self.config.type,
-                self.config.license,
-                self.config.tags,
-                cfg=cfg,
-                dataset_meta=dataset_meta,
+        warnings.warn(
+            "PreTrainedPolicy.push_model_to_hub is deprecated and will be removed in a future "
+            "version. Use lerobot.common.train_utils.publish_trained_model(cfg, model, "
+            "preprocessor, postprocessor, dataset_meta) instead.",
+            FutureWarning,
+            stacklevel=2,
+        )
+        if state_dict is not None:
+            warnings.warn(
+                "The `state_dict` argument is ignored: sharded weights are gathered internally "
+                "when the policy is saved.",
+                FutureWarning,
+                stacklevel=2,
            )
-            card.save(str(saved_path / "README.md"))
-
-            cfg.save_pretrained(saved_path)  # Calls _save_pretrained and stores train config
-
-            commit_info = api.upload_folder(
-                repo_id=repo_id,
-                repo_type="model",
-                folder_path=saved_path,
-                commit_message="Upload policy weights, train config and readme",
-                allow_patterns=["*.safetensors", "*.json", "*.yaml", "*.md"],
-                ignore_patterns=["*.tmp", "*.log"],
-            )
-
-            # Contract: lerobot.jobs.hf.submit_to_hf watches for this exact
-            # "Model pushed to <url>" line to end a remote run early. Keep the wording
-            # and URL format in sync (it falls back to status polling if they drift).
-            logging.info(f"Model pushed to {commit_info.repo_url.url}")
-
-    def generate_model_card(
-        self,
-        dataset_repo_id: str,
-        model_type: str,
-        license: str | None,
-        tags: list[str] | None,
-        cfg: TrainPipelineConfig | None = None,
-        dataset_meta: LeRobotDatasetMetadata | None = None,
-    ) -> ModelCard:
-        base_model_mapping = {
-            "smolvla": "lerobot/smolvla_base",
-            "pi0": "lerobot/pi0_base",
-            "pi05": "lerobot/pi05_base",
-            "pi0_fast": "lerobot/pi0fast-base",
-            "xvla": "lerobot/xvla-base",
-        }
-
-        card_data = ModelCardData(
-            license=license or "apache-2.0",
-            library_name="lerobot",
-            pipeline_tag="robotics",
-            tags=list(set(tags or []).union({"robotics", "lerobot", model_type})),
-            model_name=model_type,
-            datasets=dataset_repo_id,
-            base_model=base_model_mapping.get(model_type),
-        )
-
-        context = _build_card_context(
-            cfg, dataset_meta, self.config.input_features, self.config.output_features
-        )
-        # Used by the template to pre-fill commands and the "Fine-tuned from" line.
-        context["policy_repo_id"] = getattr(self.config, "repo_id", None)
-        context["base_model"] = base_model_mapping.get(model_type)
-
-        template_card = (
-            files("lerobot.templates").joinpath("lerobot_modelcard_template.md").read_text(encoding="utf-8")
-        )
-        card = ModelCard.from_template(card_data, template_str=template_card, **context)
-        card.validate()
-        return card
+        publish_trained_model(cfg, self, None, None, dataset_meta, peft_model=peft_model)

    def wrap_with_peft(
        self,
@@ -236,21 +236,11 @@ class ActionQueue:
        if action_index_before_inference is not None:
            indexes_diff = max(0, self.last_index - action_index_before_inference)
            if indexes_diff != real_delay:
-                # The latency estimate (`real_delay`) and the number of actions the robot
-                # actually consumed during inference (`indexes_diff`) disagree. This happens
-                # when the queue starved (robot idle) or on the first chunk (nothing consumed
-                # yet). Discarding `real_delay` here would drop actions the arm never executed
-                # and splice the queue `real_delay` steps ahead of the physical pose — a hard
-                # jump/slam, worst on slow policies where `real_delay` is large. Never discard
-                # more than was actually consumed.
-                resolved = min(real_delay, indexes_diff)
                logger.warning(
-                    "Indexes diff != real delay (indexes_diff=%d, real_delay=%d); "
-                    "clamping discard to %d to avoid a queue-splice jump.",
+                    "Indexes diff is not equal to real delay. indexes_diff=%d, real_delay=%d",
                    indexes_diff,
                    real_delay,
-                    resolved,
                )
-                return resolved
+                return real_delay

        return effective_delay
@@ -189,20 +189,15 @@ class DiT(ModelMixin, ConfigMixin):

@dataclass
 class ActionModelPreset:
-    """Default head geometry per `action_model_type`.
-
-    Only the attention geometry is preset; the DiT's width comes from
-    `config.action_hidden_size`, so there is deliberately no `hidden_size` here.
-    """
-
+    hidden_size: int
    attention_head_dim: int
    num_attention_heads: int


 DIT_PRESETS = {
-    "DiT-B": ActionModelPreset(attention_head_dim=64, num_attention_heads=12),
-    "DiT-L": ActionModelPreset(attention_head_dim=48, num_attention_heads=32),
-    "DiT-test": ActionModelPreset(attention_head_dim=8, num_attention_heads=2),
+    "DiT-B": ActionModelPreset(hidden_size=768, attention_head_dim=64, num_attention_heads=12),
+    "DiT-L": ActionModelPreset(hidden_size=1536, attention_head_dim=48, num_attention_heads=32),
+    "DiT-test": ActionModelPreset(hidden_size=16, attention_head_dim=8, num_attention_heads=2),
 }


@@ -253,10 +248,7 @@ class VLAJEPAActionHead(nn.Module):
        )
        self.future_tokens = nn.Embedding(config.num_embodied_action_tokens_per_instruction, inner_dim)
        self.position_embedding = nn.Embedding(
-            max(
-                config.action_max_seq_len,
-                config.chunk_size + config.num_action_tokens_per_timestep + 4,
-            ),
+            max(1024, config.chunk_size + config.num_action_tokens_per_timestep + 4),
            inner_dim,
        )
        self.beta_dist = Beta(config.action_noise_beta_alpha, config.action_noise_beta_beta)
@@ -267,15 +259,11 @@ class VLAJEPAActionHead(nn.Module):

    def _build_inputs(
        self,
+        conditioning_tokens: torch.Tensor,
        actions: torch.Tensor,
        state: torch.Tensor | None,
        timesteps: torch.Tensor,
    ) -> torch.Tensor:
-        """Build the DiT's own token sequence: [state?, future queries, noisy actions].
-
-        The conditioning tokens are not part of this sequence; they reach the DiT as
-        `encoder_hidden_states` through cross-attention.
-        """
        action_features = self.action_encoder(actions, timesteps)
        pos_ids = torch.arange(action_features.shape[1], device=actions.device)
        action_features = action_features + self.position_embedding(pos_ids)[None]
@@ -294,7 +282,6 @@ class VLAJEPAActionHead(nn.Module):
        actions: torch.Tensor,
        state: torch.Tensor | None = None,
        action_is_pad: torch.Tensor | None = None,
-        reduction: str = "mean",
    ) -> torch.Tensor:
        noise = torch.randn_like(actions)
        t = self.sample_time(actions.shape[0], actions.device, actions.dtype)
@@ -302,7 +289,7 @@ class VLAJEPAActionHead(nn.Module):
        velocity = actions - noise
        t_discretized = (t * self.config.action_num_timestep_buckets).long()

-        hidden_states = self._build_inputs(noisy_actions, state, t_discretized)
+        hidden_states = self._build_inputs(conditioning_tokens, noisy_actions, state, t_discretized)
        pred = self.model(
            hidden_states=hidden_states,
            encoder_hidden_states=conditioning_tokens,
@@ -315,10 +302,6 @@ class VLAJEPAActionHead(nn.Module):

        loss = F.mse_loss(pred_actions, velocity, reduction="none")  # [B, T, action_dim]
        valid_mask = ~action_is_pad.unsqueeze(-1)  # [B, T, 1]
-        if reduction == "none":
-            # Per-sample loss (B,) for sample weighting (RA-BC): mask-average over T and action_dim.
-            per_sample_valid = valid_mask.sum(dim=(1, 2)) * loss.shape[-1]  # [B]
-            return (loss * valid_mask).sum(dim=(1, 2)) / per_sample_valid.clamp_min(1)
        num_valid = valid_mask.sum() * loss.shape[-1]
        return (loss * valid_mask).sum() / num_valid.clamp_min(1)

@@ -343,7 +326,7 @@ class VLAJEPAActionHead(nn.Module):
            timesteps = torch.full(
                (batch_size,), t_value, device=conditioning_tokens.device, dtype=torch.long
            )
-            hidden_states = self._build_inputs(actions, state, timesteps)
+            hidden_states = self._build_inputs(conditioning_tokens, actions, state, timesteps)
            pred = self.model(
                hidden_states=hidden_states,
                encoder_hidden_states=conditioning_tokens,
@@ -58,14 +58,6 @@ class VLAJEPAConfig(PreTrainedConfig):
    action_dim: int = 7
    state_dim: int = 8

-    # Relative actions: converts absolute actions to relative (action -= state) during
-    # preprocessing, and reverses it at postprocessing. Requires `state_dim` (OBS_STATE).
-    use_relative_actions: bool = False
-    # Joint names to keep absolute (not converted to relative). Empty list = all dims relative.
-    relative_exclude_joints: list[str] = field(default_factory=lambda: ["gripper"])
-    # Populated at runtime from dataset metadata by make_policy (used to build the exclude mask).
-    action_feature_names: list[str] | None = None
-
    num_action_tokens_per_timestep: int = 8
    num_embodied_action_tokens_per_instruction: int = 32
    num_inference_timesteps: int = 4
@@ -80,12 +72,8 @@ class VLAJEPAConfig(PreTrainedConfig):
    action_noise_beta_alpha: float = 1.5
    action_noise_beta_beta: float = 1.0
    action_noise_s: float = 0.999
-    # Size of the action head's learned position-embedding table. Kept at 1024 to match the
-    # published checkpoints; only raise it if `chunk_size` approaches that.
-    action_max_seq_len: int = 1024
-    # Unused. Retained because the published checkpoints serialize it and draccus rejects
-    # config.json keys that the dataclass no longer declares.
    num_target_vision_tokens: int = 32
+    action_max_seq_len: int = 1024

    # total video frames loaded per sample
    num_video_frames: int = 8
@@ -94,34 +82,15 @@ class VLAJEPAConfig(PreTrainedConfig):
    predictor_mlp_ratio: float = 4.0
    predictor_dropout: float = 0.0
    world_model_loss_weight: float = 0.1
-    # Temporal tubelet size of the JEPA encoder (e.g. 2 for vjepa2-vitl-fpc64-256). When the
-    # world model is enabled the encoder's own `config.tubelet_size` is authoritative and this
-    # is only used for the `num_video_frames` sanity check below.
-    jepa_tubelet_size: int = 2
-    # Number of camera views the world model's predictor is built for; its embedding width is
-    # `encoder_hidden_size * world_model_num_views`, so it is baked into the checkpoint shapes.
-    # Views beyond this are trimmed and missing ones are padded with the first view.
-    # `None` falls back to `jepa_tubelet_size`, which is what the published checkpoints
-    # (trained before the two meanings were separated) actually encode.
-    world_model_num_views: int | None = None
+    jepa_tubelet_size: int = 2  # must match the encoder (e.g. 2 for vjepa2-vitl-fpc64-256)
    repeated_diffusion_steps: int = 8  # independent noise draws per batch item (CogACT-style)

    resize_images_to: tuple[int, int] | None = None
-    # Gripper post-processing, ported from the starVLA LIBERO eval loop. OFF by default
-    # because it is only correct for LIBERO's action convention: `pre_snap` writes {0, 1}
-    # into normalized space and `binarize` then thresholds the *unnormalized* value at
-    # `gripper_threshold`, so a gripper whose physical range is not roughly [0, 1] (degrees,
-    # mm, [0, 100]) gets pinned to a constant. Enable them only for LIBERO-style setups.
-    binarize_gripper_action: bool = False
-    pre_snap_gripper_action: bool = False
+    binarize_gripper_action: bool = True
+    pre_snap_gripper_action: bool = True
    clip_normalized_actions: bool = True
-    # Index of the gripper in the action vector. Prefer leaving this at its default and
-    # setting `gripper_joint_names`, which resolves the index from dataset metadata.
    gripper_dim: int = 6
    gripper_threshold: float = 0.5
-    # Action-dimension names identifying the gripper. When these match `action_feature_names`,
-    # the resolved index wins over `gripper_dim`.
-    gripper_joint_names: list[str] = field(default_factory=lambda: ["gripper"])
    torch_dtype: str = "bfloat16"

    optimizer_lr: float = 1e-4
@@ -146,27 +115,6 @@ class VLAJEPAConfig(PreTrainedConfig):
                f"({self.jepa_tubelet_size}) to have at least one context and one GT temporal position."
            )

-    @property
-    def num_world_model_views(self) -> int:
-        """Camera views the world model predictor is built for (see `world_model_num_views`)."""
-        return self.world_model_num_views or self.jepa_tubelet_size
-
-    @property
-    def resolved_gripper_dim(self) -> int:
-        """Gripper index, resolved from `action_feature_names` when possible.
-
-        Falls back to the raw `gripper_dim` when dataset metadata is unavailable (for example
-        when a saved processor pipeline is rebuilt without a dataset attached).
-        """
-        if not self.action_feature_names or not self.gripper_joint_names:
-            return self.gripper_dim
-        wanted = [name.lower() for name in self.gripper_joint_names if name]
-        for index, name in enumerate(self.action_feature_names):
-            lowered = str(name).lower()
-            if any(token == lowered or token in lowered for token in wanted):
-                return index
-        return self.gripper_dim
-
    def validate_features(self) -> None:
        if not self.image_features:
            raise ValueError("VLAJEPA requires at least one visual input feature.")
@@ -175,16 +123,6 @@ class VLAJEPAConfig(PreTrainedConfig):
        self.action_dim = self.action_feature.shape[0]
        if self.robot_state_feature is not None:
            self.state_dim = self.robot_state_feature.shape[0]
-        # The gripper steps silently no-op when the index is out of range, which reads as
-        # "binarization ran" while nothing happened. Fail loudly at construction instead.
-        if self.pre_snap_gripper_action or self.binarize_gripper_action:
-            gripper_dim = self.resolved_gripper_dim
-            if gripper_dim >= self.action_dim:
-                raise ValueError(
-                    f"`gripper_dim` ({gripper_dim}) is out of range for a {self.action_dim}-dim "
-                    f"action. Set `gripper_dim`/`gripper_joint_names` to the real gripper index, "
-                    f"or disable `pre_snap_gripper_action`/`binarize_gripper_action`."
-                )

    def set_dataset_feature_metadata(self, dataset_features: dict[str, Any]) -> None:
        """Add `observation.state` to `input_features` if missing, so it gets normalized."""
@@ -212,18 +150,9 @@ class VLAJEPAConfig(PreTrainedConfig):

    @property
    def observation_delta_indices(self) -> list[int]:
-        # Only the world model consumes frames past index 0, so without it asking for the full
-        # window would decode `num_video_frames` frames per camera per sample and drop them.
-        if not self.enable_world_model:
-            return [0]
-        # matches original repo's observation_indices=list(range(video_horizon)) when the chunk
-        # fits within video_horizon frames. When chunk_size is longer (e.g. folding's 30-step
-        # chunk vs 8 video frames), spread the frames evenly across the chunk instead of
-        # clustering them at the start, so the world model sees dynamics over the whole horizon.
-        if self.num_video_frames >= self.chunk_size:
-            return list(range(self.num_video_frames))
-        stride = (self.chunk_size - 1) // (self.num_video_frames - 1)
-        return [i * stride for i in range(self.num_video_frames)]
+        # load video_horizon frames starting from current timestep: [t, t+1, ..., t+video_horizon-1]
+        # matches original repo's observation_indices=list(range(video_horizon))
+        return list(range(self.num_video_frames))

    @property
    def action_delta_indices(self) -> list[int]:
@@ -17,17 +17,17 @@ from __future__ import annotations
 import logging
 from collections import deque
 from contextlib import nullcontext
+from pathlib import Path
 from typing import TYPE_CHECKING, Any

 import torch
 import torch.nn.functional as F  # noqa: N812
-from safetensors.torch import load_file
 from torch import Tensor, nn

 from lerobot.policies.pretrained import PreTrainedPolicy, T
-from lerobot.policies.utils import log_model_loading_keys, populate_queues
+from lerobot.policies.utils import populate_queues
 from lerobot.utils.constants import ACTION, OBS_STATE
-from lerobot.utils.device_utils import is_amp_available, resolve_safetensors_device
+from lerobot.utils.device_utils import is_amp_available
 from lerobot.utils.import_utils import _transformers_available, require_package

 if TYPE_CHECKING or _transformers_available:
@@ -51,11 +51,6 @@ def _get_autocast_context(device_type: str, dtype: torch.dtype = torch.bfloat16)
    """
    if not is_amp_available(device_type):
        return nullcontext()
-    if device_type == "cpu" and dtype not in (torch.bfloat16, torch.float16):
-        # CPU autocast does not implement float32 (used here to force the action head back to
-        # full precision). `torch.autocast` accepts it, then warns and disables itself on *every*
-        # call, so short-circuit to keep the same behavior without the per-forward log spam.
-        return nullcontext()
    if device_type == "cuda" and dtype == torch.bfloat16 and not torch.cuda.is_bf16_supported():
        dtype = torch.float16
    return torch.autocast(device_type=device_type, dtype=dtype)
@@ -112,7 +107,7 @@ class VLAJEPAModel(nn.Module):
                torch_dtype=self.qwen._get_torch_dtype(config.torch_dtype),
            )
            self.video_processor = AutoVideoProcessor.from_pretrained(config.jepa_encoder_name)
-            num_views = config.num_world_model_views
+            num_views = config.jepa_tubelet_size
            tubelet_size = self.video_encoder.config.tubelet_size
            image_size = getattr(self.video_encoder.config, "image_size", None)
            if image_size is None:
@@ -130,7 +125,6 @@ class VLAJEPAModel(nn.Module):
                num_heads=config.predictor_num_heads,
                mlp_ratio=config.predictor_mlp_ratio,
                num_action_tokens_per_step=config.num_action_tokens_per_timestep,
-                dropout=config.predictor_dropout,
            )
        else:
            self.video_encoder = None
@@ -165,13 +159,6 @@ class VLAJEPAModel(nn.Module):
        `output_hidden_states=True` is post-norm (tied to `last_hidden_state` via
        `@capture_outputs`). A forward hook on `language_model.layers[-1]` recovers
        the correct pre-RMSNorm state, matching the training-time representation.
-
-        Calls the inner `Qwen3VLModel` rather than the `Qwen3VLForConditionalGeneration`
-        wrapper: only the hooked hidden state is used, and the wrapper's forward ends in
-        `lm_head(hidden_states[:, slice(None), :])` because `logits_to_keep` defaults to 0,
-        so it would build full-sequence logits over the 151936-token vocab and discard them
-        (~3.4 GB in bf16 at batch 8). The wrapper stays as `self.qwen.model` so `lm_head`
-        keeps its checkpoint key; only this forward path skips it.
        """
        captured: list[torch.Tensor] = []

@@ -182,7 +169,12 @@ class VLAJEPAModel(nn.Module):
        last_layer = self.qwen.model.model.language_model.layers[-1]
        handle = last_layer.register_forward_hook(_hook)
        try:
-            self.qwen.model.model(**qwen_inputs)
+            self.qwen.model(
+                **qwen_inputs,
+                output_hidden_states=False,
+                output_attentions=False,
+                return_dict=True,
+            )
        finally:
            handle.remove()

@@ -192,7 +184,7 @@ class VLAJEPAModel(nn.Module):

    def _encode_qwen(
        self, images: list[list[Tensor]], instructions: list[str], *, need_action_tokens: bool
-    ) -> tuple[Tensor, Tensor | None]:
+    ) -> tuple[Tensor, Tensor, Tensor | None]:
        """Run Qwen and gather the embodied-action (and optionally action) token hidden states."""
        qwen_inputs = self.qwen.build_inputs(
            images=images,
@@ -219,14 +211,10 @@ class VLAJEPAModel(nn.Module):
            )
        return embodied_action_tokens, action_tokens

-    def _world_model_loss(self, videos: Tensor, action_tokens: Tensor, reduction: str = "mean") -> Tensor:
-        """JEPA encode + predictor L1 loss. `videos` is [B, V, T, C, H, W] float in [0, 1].
-
-        `reduction="none"` returns a per-sample loss (B,) for sample weighting (RA-BC);
-        "mean" returns the scalar loss.
-        """
+    def _world_model_loss(self, videos: Tensor, action_tokens: Tensor) -> Tensor:
+        """JEPA encode + predictor L1 loss. `videos` is [B, V, T, C, H, W] float in [0, 1]."""
        # Match the world model's expected view count: pad with the first view, or trim extras.
-        num_views = self.config.num_world_model_views
+        num_views = self.config.jepa_tubelet_size
        if videos.shape[1] < num_views:
            missing = num_views - videos.shape[1]
            videos = torch.cat([videos, videos[:, :1].repeat(1, missing, 1, 1, 1, 1)], dim=1)
@@ -245,24 +233,14 @@ class VLAJEPAModel(nn.Module):

        with torch.no_grad():
            video_embeddings = self.video_encoder.get_vision_features(pixel_values_videos=video_pixels)
-            # Merge views: [B*V, N, H] -> [B, N, V*H].
-            # `flat` above flattens (B, V) row-major, so rows run view-fastest:
-            # (s0v0, s0v1, ..., s1v0, ...). A `chunk(chunks=v, dim=0)` + `cat(dim=2)` would
-            # instead assume view-slowest ordering and concatenate features belonging to
-            # *different samples* — shape-valid, so it fails silently. Regroup on (B, V).
-            n_tokens, hidden = video_embeddings.shape[1], video_embeddings.shape[2]
-            video_embeddings = (
-                video_embeddings.view(b, v, n_tokens, hidden)
-                .permute(0, 2, 1, 3)
-                .reshape(b, n_tokens, v * hidden)
-            )
+            # Merge views: [B*V, ...] -> [B, ..., V*embed_dim]
+            video_embeddings = torch.cat(torch.chunk(video_embeddings, chunks=v, dim=0), dim=2)

        tubelet_size = self.video_encoder.config.tubelet_size
        # num_video_frames raw frames → t_enc_total temporal positions after tubelet compression
        t_enc_total = self.config.num_video_frames // tubelet_size
        if t_enc_total < 2:
-            zero_shape = (video_embeddings.shape[0],) if reduction == "none" else ()
-            return torch.zeros(zero_shape, device=video_embeddings.device)
+            return torch.zeros((), device=video_embeddings.device)

        # Shift-by-one JEPA split: input_states = positions 0..T-2, gt_states = positions 1..T-1
        t_enc_ctx = t_enc_total - 1
@@ -278,10 +256,6 @@ class VLAJEPAModel(nn.Module):
        predicted_states = self.video_predictor(
            input_states.float(), action_tokens[:, :expected_actions].float()
        )
-        if reduction == "none":
-            # Per-sample loss (B,): mean over all non-batch dims (tokens, feature).
-            l = F.l1_loss(predicted_states, gt_states.float(), reduction="none")
-            return l.mean(dim=tuple(range(1, l.ndim)))
        return F.l1_loss(predicted_states, gt_states.float(), reduction="mean")

    def _action_loss(
@@ -290,27 +264,17 @@ class VLAJEPAModel(nn.Module):
        actions: Tensor,
        state: Tensor | None,
        action_is_pad: Tensor | None,
-        reduction: str = "mean",
    ) -> Tensor:
-        """Flow-matching action-head loss, repeated over `repeated_diffusion_steps`.
-
-        `reduction="none"` returns a per-sample loss (B,) — the `repeated_diffusion_steps`
-        independent noise draws are averaged back per original sample — for RA-BC weighting.
-        """
+        """Flow-matching action-head loss, repeated over `repeated_diffusion_steps`."""
        device_type = next(self.parameters()).device.type
        with _get_autocast_context(device_type, torch.float32):
            r = self.config.repeated_diffusion_steps
            horizon = self.config.chunk_size
-            b = embodied_action_tokens.shape[0]
            actions_target = actions[:, -horizon:, :].to(torch.float32).repeat(r, 1, 1)
            embodied = embodied_action_tokens.repeat(r, 1, 1)
            state_rep = state.to(embodied_action_tokens.dtype).repeat(r, 1, 1) if state is not None else None
            pad_rep = action_is_pad[:, -horizon:].repeat(r, 1) if action_is_pad is not None else None
-            loss = self.action_model(embodied, actions_target, state_rep, pad_rep, reduction=reduction)
-            if reduction == "none":
-                # `.repeat(r, 1, 1)` tiles as [rep0(b0..b_{B-1}), rep1(...), ...] → (r, B); mean over reps.
-                return loss.view(r, b).mean(dim=0)
-            return loss
+            return self.action_model(embodied, actions_target, state_rep, pad_rep)

    def forward(
        self,
@@ -320,29 +284,21 @@ class VLAJEPAModel(nn.Module):
        actions: Tensor | None = None,
        state: Tensor | None = None,
        action_is_pad: Tensor | None = None,
-        reduction: str = "mean",
    ) -> dict[str, Tensor]:
-        """Native forward: Qwen encode → optional world-model loss → optional action-head loss.
-
-        `reduction="none"` makes both loss terms per-sample (B,) for RA-BC weighting; "mean"
-        returns scalar losses.
-        """
+        """Native forward: Qwen encode → optional world-model loss → optional action-head loss."""
        embodied_action_tokens, action_tokens = self._encode_qwen(
            images, instructions, need_action_tokens=self.config.enable_world_model
        )

        if self.config.enable_world_model and videos is not None:
-            wm_loss = self._world_model_loss(videos, action_tokens, reduction=reduction)
+            wm_loss = self._world_model_loss(videos, action_tokens)
        else:
-            zero_shape = (embodied_action_tokens.shape[0],) if reduction == "none" else ()
-            wm_loss = torch.zeros(zero_shape, device=embodied_action_tokens.device)
+            wm_loss = torch.zeros((), device=embodied_action_tokens.device)

        if actions is None:
            return {"wm_loss": wm_loss}

-        action_loss = self._action_loss(
-            embodied_action_tokens, actions, state, action_is_pad, reduction=reduction
-        )
+        action_loss = self._action_loss(embodied_action_tokens, actions, state, action_is_pad)
        return {"action_loss": action_loss, "wm_loss": wm_loss * self.config.world_model_loss_weight}

    # ---- Native predict_action (follows original VLA_JEPA.predict_action) ----
@@ -428,19 +384,12 @@ class VLAJEPAPolicy(PreTrainedPolicy):
        batch_size = batch[image_keys[0]].shape[0]

        # Current-frame image per view ([B, C, H, W]); regroup per sample for Qwen messages.
-        # Resize to config.resize_images_to (as predict_action does) so training and inference feed
-        # Qwen the same resolution. Critical for memory: native camera frames (e.g. 720x1280) would
-        # otherwise blow up the Qwen3-VL vision-tower attention (patch count grows with resolution).
-        resize_hw = tuple(self.config.resize_images_to) if self.config.resize_images_to else None
        frames = []
        for key in image_keys:
            t = batch[key]
            if t.ndim == 5:  # [B, T, C, H, W] -> current observation (delta=0)
                t = t[:, 0]
-            px = self.model.qwen.to_pixel_values(t)  # [B, C, H, W]
-            if resize_hw is not None and tuple(px.shape[-2:]) != resize_hw:
-                px = F.interpolate(px.float(), size=resize_hw, mode="area")
-            frames.append(px)
+            frames.append(self.model.qwen.to_pixel_values(t))
        images = [[frame[b] for frame in frames] for b in range(batch_size)]

        tasks = batch.get("task")
@@ -456,26 +405,7 @@ class VLAJEPAPolicy(PreTrainedPolicy):
        # Videos [B, V, T, C, H, W] - only assembled during training when the world model consumes them.
        if self.model.config.enable_world_model and training:
            views = [batch[k].unsqueeze(1) if batch[k].ndim == 4 else batch[k] for k in image_keys]
-            # The world model consumes a SINGLE stacked [B, V, T, C, H, W] tensor, so all camera
-            # views must share a spatial size. Cameras can differ (e.g. base 480x640 vs wrist
-            # 720x1280), so resize each view to a common size before stacking — config.resize_images_to
-            # if set (same target predict_action uses), else the first view's size (a no-op when all
-            # views already match, preserving behavior for single-resolution datasets). The vjepa video
-            # processor does the final resize to the encoder resolution downstream.
-            cfg = self.model.config
-            target_hw = tuple(cfg.resize_images_to) if cfg.resize_images_to else tuple(views[0].shape[-2:])
-            resized = []
-            for v in views:
-                if tuple(v.shape[-2:]) != target_hw:
-                    b, t, c = v.shape[0], v.shape[1], v.shape[2]
-                    v = F.interpolate(
-                        v.reshape(b * t, c, v.shape[3], v.shape[4]).float(),
-                        size=target_hw,
-                        mode="bilinear",
-                        align_corners=False,
-                    ).reshape(b, t, c, target_hw[0], target_hw[1])
-                resized.append(v)
-            inputs["videos"] = self.model.qwen.to_pixel_values(torch.stack(resized, dim=1))
+            inputs["videos"] = self.model.qwen.to_pixel_values(torch.stack(views, dim=1))

        actions = batch.get(ACTION)
        if actions is not None:
@@ -494,17 +424,15 @@ class VLAJEPAPolicy(PreTrainedPolicy):

    # ---- LeRobot Policy Interface ----

-    def forward(self, batch: dict[str, Tensor], reduction: str = "mean") -> tuple[Tensor, dict]:
+    def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict]:
        """LeRobot train forward: convert → native forward → aggregate losses."""
-        native_output = self.model.forward(
-            **self._prepare_model_inputs(batch, training=True), reduction=reduction
-        )
+        native_output = self.model.forward(**self._prepare_model_inputs(batch, training=True))

        ref = next(iter(native_output.values()))
-        zero = torch.zeros_like(ref)
+        zero = torch.zeros((), device=ref.device, dtype=ref.dtype)
        total_loss = native_output.get("action_loss", zero) + native_output.get("wm_loss", zero)
-        logs = {k: v.detach().mean().item() for k, v in native_output.items()}
-        logs["loss"] = total_loss.detach().mean().item()
+        logs = {k: v.detach().item() for k, v in native_output.items()}
+        logs["loss"] = total_loss.detach().item()
        return total_loss, logs

    def get_optim_params(self) -> dict:
@@ -530,16 +458,23 @@ class VLAJEPAPolicy(PreTrainedPolicy):
            self._queues[ACTION].extend(actions.transpose(0, 1)[: self.config.n_action_steps])
        return self._queues[ACTION].popleft()

+    @classmethod
+    def from_pretrained(
+        cls: type[T],
+        pretrained_name_or_path: str | Path,
+        **kwargs,
+    ):
+        return super().from_pretrained(pretrained_name_or_path, **kwargs)
+
    @classmethod
    def _load_as_safetensor(cls, model: T, model_file: str, map_location: str, strict: bool) -> T:
        reinit_prefixes = model.config.reinit_modules
        if not reinit_prefixes:
            return super()._load_as_safetensor(model, model_file, map_location, strict)

-        # `resolve_safetensors_device` is what keeps every rank from materializing the whole
-        # checkpoint on GPU 0: safetensors maps the bare string "cuda" to cuda:0 regardless of
-        # torch.cuda.current_device(), and `config.device` is exactly that bare string.
-        state_dict = load_file(model_file, device=resolve_safetensors_device(map_location))
+        from safetensors.torch import load_file
+
+        state_dict = load_file(model_file, device=map_location)
        current = model.state_dict()

        reinitialized: list[str] = []
@@ -563,17 +498,8 @@ class VLAJEPAPolicy(PreTrainedPolicy):
                f"(randomly re-initialised):\n  " + "\n  ".join(reinitialized)
            )

-        # Deliberately non-strict: the reinitialized tensors above are *expected* to be missing.
-        # `strict` still has to mean something, so enforce it on everything else.
+        from lerobot.policies.utils import log_model_loading_keys
+
        missing_keys, unexpected_keys = model.load_state_dict(filtered, strict=False)
-        if strict:
-            reinit_keys = {entry.split(":", 1)[0] for entry in reinitialized}
-            unaccounted = [k for k in missing_keys if k not in reinit_keys]
-            if unaccounted or unexpected_keys:
-                raise RuntimeError(
-                    f"Error(s) in loading state_dict for {type(model).__name__} with strict=True: "
-                    f"missing keys not covered by `reinit_modules` {unaccounted}, "
-                    f"unexpected keys {list(unexpected_keys)}."
-                )
        log_model_loading_keys(missing_keys, unexpected_keys)
        return model
@@ -14,93 +14,22 @@

 from __future__ import annotations

-import logging
 from typing import Any

 import torch
-import torch.nn.functional as F  # noqa: N812

-from lerobot.configs import PipelineFeatureType, PolicyFeature
-from lerobot.configs.types import NormalizationMode
 from lerobot.policies.vla_jepa.configuration_vla_jepa import VLAJEPAConfig
 from lerobot.processor import (
-    AbsoluteActionsProcessorStep,
    EnvTransition,
-    ObservationProcessorStep,
    PolicyAction,
    PolicyProcessorPipeline,
    ProcessorStep,
    ProcessorStepRegistry,
-    RelativeActionsProcessorStep,
    TransitionKey,
    UnnormalizerProcessorStep,
    make_default_policy_processor_steps,
    make_policy_processor_pipelines,
 )
-from lerobot.utils.constants import ACTION
-
-
-@ProcessorStepRegistry.register(name="vla_jepa_image_prep")
-class ImagePrepProcessorStep(ObservationProcessorStep):
-    """Prepares image observations for the VLA-JEPA model: float cast, 1->3 channel expand, resize.
-
-    This makes explicit (in the serialized pipeline) the image prep the model used to do
-    internally. The model keeps the same operations as idempotent guards, so:
-      - checkpoints saved WITHOUT this step (older uploads) are unaffected — the model still
-        does the prep;
-      - checkpoints saved WITH this step get it done here, and the model-side guards no-op.
-
-    Mirrors `Qwen3VLInterface.to_pixel_values` + the `F.interpolate(mode="area")` resize in
-    `VLAJEPAPolicy._prepare_model_inputs`/`predict_action`. Deliberately does NOT clamp (the
-    model path doesn't), so values stay bit-identical. Handles [C,H,W], [B,C,H,W]/[T,C,H,W]
-    and [B,T,C,H,W] image tensors.
-    """
-
-    def __init__(self, resize_to: tuple[int, int] | None = None, expand_channels: bool = True):
-        self.resize_to = tuple(resize_to) if resize_to is not None else None
-        self.expand_channels = expand_channels
-
-    def observation(self, observation: dict) -> dict:
-        new_observation = dict(observation)
-        for key in observation:
-            if "image" not in key:
-                continue
-            image = observation[key].float()
-            if self.expand_channels and image.shape[-3] == 1:
-                repeats = [1] * image.ndim
-                repeats[-3] = 3
-                image = image.repeat(*repeats)
-            if self.resize_to is not None and tuple(image.shape[-2:]) != self.resize_to:
-                device = image.device
-                # NOTE: no "area" kernel on mps; resize on cpu then move back.
-                if device.type == "mps":
-                    image = image.cpu()
-                lead = image.shape[:-3]
-                c, h, w = image.shape[-3:]
-                flat = image.reshape(-1, c, h, w)
-                flat = F.interpolate(flat, size=self.resize_to, mode="area")
-                image = flat.reshape(*lead, c, *self.resize_to).to(device)
-            new_observation[key] = image
-        return new_observation
-
-    def get_config(self) -> dict[str, Any]:
-        return {
-            "resize_to": list(self.resize_to) if self.resize_to is not None else None,
-            "expand_channels": self.expand_channels,
-        }
-
-    def transform_features(self, features):
-        for key in features[PipelineFeatureType.OBSERVATION]:
-            if "image" not in key:
-                continue
-            feat = features[PipelineFeatureType.OBSERVATION][key]
-            # Match `to_pixel_values`: only a single channel is expanded to 3.
-            nb_channel = 3 if (self.expand_channels and feat.shape[0] == 1) else feat.shape[0]
-            spatial = self.resize_to if self.resize_to is not None else tuple(feat.shape[1:])
-            features[PipelineFeatureType.OBSERVATION][key] = PolicyFeature(
-                type=feat.type, shape=(nb_channel, *spatial)
-            )
-        return features


@ProcessorStepRegistry.register(name="vla_jepa_clip_actions")
@@ -142,11 +71,6 @@ class PreSnapGripperProcessorStep(ProcessorStep):
            transition[TransitionKey.ACTION] = a
        return transition

-    def get_config(self) -> dict[str, Any]:
-        # Without this the base class serializes `{}` and a reloaded pipeline silently reverts
-        # to the class defaults, discarding whatever the training config set.
-        return {"gripper_dim": self.gripper_dim, "threshold": self.threshold}
-
    def transform_features(self, features):
        return features

@@ -157,13 +81,6 @@ class BinarizeGripperProcessorStep(ProcessorStep):

    Maps continuous value to {-1, 1}: > threshold → -1, <= threshold → 1 (matches starVLA convention).
    Only applied when action has more dimensions than gripper_dim.
-
-    WARNING: this step runs *below* the unnormalizer, so `threshold` is compared against the
-    gripper's **physical** value while its default of 0.5 comes from the model's [0, 1]/±1
-    convention. It is only meaningful when the gripper's physical range is roughly [0, 1].
-    For a gripper in degrees, mm or [0, 100], every unnormalized value exceeds 0.5 and the
-    output collapses to the constant -1. `make_vla_jepa_pre_post_processors` warns when the
-    dataset stats say that is the case.
    """

    def __init__(self, gripper_dim: int = 6, threshold: float = 0.5):
@@ -179,57 +96,10 @@ class BinarizeGripperProcessorStep(ProcessorStep):
            transition[TransitionKey.ACTION] = a
        return transition

-    def get_config(self) -> dict[str, Any]:
-        # See PreSnapGripperProcessorStep.get_config: `{}` would reload as the class defaults.
-        return {"gripper_dim": self.gripper_dim, "threshold": self.threshold}
-
    def transform_features(self, features):
        return features


-def _warn_if_gripper_steps_are_misconfigured(
-    config: VLAJEPAConfig,
-    gripper_dim: int,
-    dataset_stats: dict[str, dict[str, torch.Tensor]] | None,
-) -> None:
-    """Warn when the gripper post-steps would pin the gripper to a constant.
-
-    `BinarizeGripperProcessorStep` thresholds the *unnormalized* gripper at
-    `gripper_threshold` (default 0.5, a number from the model's [0, 1] convention). When the
-    dataset says the gripper's physical range sits well above that, every unnormalized value
-    lands on the same side of the threshold and the commanded gripper never moves. The stats
-    needed to detect that are already here, so say so rather than letting it look like it worked.
-    """
-    if not (config.pre_snap_gripper_action or config.binarize_gripper_action):
-        return
-    action_stats = (dataset_stats or {}).get(ACTION)
-    if not action_stats or "min" not in action_stats or "max" not in action_stats:
-        return
-    try:
-        low = float(action_stats["min"][gripper_dim])
-        high = float(action_stats["max"][gripper_dim])
-    except (IndexError, TypeError, ValueError):
-        return
-    threshold = config.gripper_threshold
-    # `pre_snap` writes {0, 1} in normalized space, which unnormalizes to the midpoint and the
-    # max. Both landing on the same side of the threshold means a constant output.
-    midpoint = (low + high) / 2.0
-    if (midpoint > threshold) == (high > threshold):
-        name = (
-            config.action_feature_names[gripper_dim]
-            if config.action_feature_names and gripper_dim < len(config.action_feature_names)
-            else f"dim {gripper_dim}"
-        )
-        logging.warning(
-            f"vla_jepa gripper post-processing looks misconfigured: action {name} has a physical "
-            f"range of [{low:.3g}, {high:.3g}], and `gripper_threshold={threshold}` is compared "
-            f"against that unnormalized value. Both {midpoint:.3g} and {high:.3g} fall on the same "
-            f"side of it, so the commanded gripper will be constant. Set `gripper_threshold` in "
-            f"the gripper's own units, or set `pre_snap_gripper_action=false` and "
-            f"`binarize_gripper_action=false` (the defaults) unless you are running LIBERO."
-        )
-
-
 def make_vla_jepa_pre_post_processors(
    config: VLAJEPAConfig,
    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
@@ -239,47 +109,18 @@ def make_vla_jepa_pre_post_processors(
 ]:
    features = {**config.input_features, **config.output_features}
    steps = make_default_policy_processor_steps(config, dataset_stats)
-
-    # Shared relative-action step (OpenPI order: raw -> relative -> normalize -> model ->
-    # unnormalize -> absolute). The SAME instance is passed to AbsoluteActionsProcessorStep
-    # below so its cached raw state (set during preprocessing) flows to postprocessing.
-    relative_step = RelativeActionsProcessorStep(
-        enabled=config.use_relative_actions,
-        exclude_joints=getattr(config, "relative_exclude_joints", []),
-        action_names=getattr(config, "action_feature_names", None),
-    )
-
    input_steps = [
        steps.rename_observations,
        steps.add_batch_dim,
        steps.to_device,
-        ImagePrepProcessorStep(
-            resize_to=tuple(config.resize_images_to) if config.resize_images_to else None,
-        ),
-        relative_step,
        steps.normalize,
    ]
-    gripper_dim = config.resolved_gripper_dim
-    _warn_if_gripper_steps_are_misconfigured(config, gripper_dim, dataset_stats)
-
    output_steps: list[ProcessorStep] = []
    if config.clip_normalized_actions:
-        # Clipping to [-1, 1] is a range assertion under MIN_MAX, but under MEAN_STD the same
-        # clamp truncates every action beyond 1 sigma. That shows up as a hesitant, low-amplitude
-        # policy with no error anywhere, so refuse to add the step instead of honoring the flag.
-        action_norm_mode = config.normalization_mapping.get("ACTION")
-        if action_norm_mode == NormalizationMode.MIN_MAX:
-            output_steps.append(ClipActionsProcessorStep())
-        else:
-            logging.warning(
-                f"`clip_normalized_actions=True` is ignored: it clips normalized actions to "
-                f"[-1, 1], which is only a no-op bound under MIN_MAX, but ACTION uses "
-                f"{getattr(action_norm_mode, 'value', action_norm_mode)}. Under MEAN_STD this "
-                f"would clamp every action to 1 sigma."
-            )
+        output_steps.append(ClipActionsProcessorStep())
    if config.pre_snap_gripper_action:
        output_steps.append(
-            PreSnapGripperProcessorStep(gripper_dim=gripper_dim, threshold=config.gripper_threshold)
+            PreSnapGripperProcessorStep(gripper_dim=config.gripper_dim, threshold=config.gripper_threshold)
        )
    # NOTE: unlike the default policy unnormalizer (output features only), VLA-JEPA
    # unnormalizes over BOTH input and output features.
@@ -290,14 +131,9 @@ def make_vla_jepa_pre_post_processors(
            stats=dataset_stats,
        )
    )
-    # Reverse the relative conversion on the unnormalized action, before gripper binarization.
-    # gripper is kept absolute by relative_exclude_joints, so the two steps touch disjoint dims.
-    output_steps.append(
-        AbsoluteActionsProcessorStep(enabled=config.use_relative_actions, relative_step=relative_step)
-    )
    if config.binarize_gripper_action:
        output_steps.append(
-            BinarizeGripperProcessorStep(gripper_dim=gripper_dim, threshold=config.gripper_threshold)
+            BinarizeGripperProcessorStep(gripper_dim=config.gripper_dim, threshold=config.gripper_threshold)
        )
    output_steps.append(steps.to_cpu)
    return make_policy_processor_pipelines(input_steps=input_steps, output_steps=output_steps)
@@ -14,7 +14,6 @@

 from __future__ import annotations

-import logging
 from collections.abc import Sequence
 from typing import TYPE_CHECKING

@@ -71,21 +70,7 @@ class Qwen3VLInterface(torch.nn.Module):
            tokenizer.add_tokens([embodied_action_token], special_tokens=True)
        embodied_action_token_id = tokenizer.convert_tokens_to_ids(embodied_action_token)

-        # Qwen3-VL-2B ships 151936 embedding rows for a 151669-token vocab, i.e. 267 spare
-        # rows, so `chunk_size * 4 + 1` added tokens fit without a resize up to chunk_size=66.
-        # Past that, resizing changes `embed_tokens` and `lm_head` shapes, and the checkpoint
-        # will no longer load against a model built with a smaller chunk_size (or vice versa)
-        # unless those prefixes are listed in `reinit_modules`. Silent until now.
-        current_rows = self.model.get_input_embeddings().weight.size(0)
-        if current_rows < len(tokenizer):
-            logging.warning(
-                f"chunk_size={self.config.chunk_size} needs {max_action_tokens + 1} added tokens, "
-                f"which exceeds the {current_rows} embedding rows of {self.config.qwen_model_name}. "
-                f"Resizing to {len(tokenizer)} rows changes the shapes of "
-                f"`model.qwen.model.model.language_model.embed_tokens` and "
-                f"`model.qwen.model.lm_head`, so this model will not load from a checkpoint trained "
-                f"with a different chunk_size unless those prefixes are in `reinit_modules`."
-            )
+        if self.model.get_input_embeddings().weight.size(0) < len(tokenizer):
            self.model.resize_token_embeddings(len(tokenizer))
        return action_tokens, action_token_ids, embodied_action_token_id

@@ -114,9 +114,9 @@ class ACRoPEAttention(nn.Module):
        self.head_dim = dim // num_heads
        self.scale = qk_scale or self.head_dim**-0.5
        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
-        self.attn_drop_prob = attn_drop
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(dim, dim)
+        self.proj_drop_prob = proj_drop
        self.proj_drop = nn.Dropout(proj_drop)
        self.use_sdpa = use_sdpa
        self.d_dim = int(2 * ((self.head_dim // 3) // 2))
@@ -231,15 +231,8 @@ class ACRoPEAttention(nn.Module):
            v = merge(v, action_v)

        if attn_mask is not None or self.use_sdpa:
-            # Attention dropout (not projection dropout), and only while training — SDPA does
-            # not consult `self.training` on its own the way `nn.Dropout` does.
            x = F.scaled_dot_product_attention(
-                q,
-                k,
-                v,
-                dropout_p=self.attn_drop_prob if self.training else 0.0,
-                is_causal=self.is_causal,
-                attn_mask=attn_mask,
+                q, k, v, dropout_p=self.proj_drop_prob, is_causal=self.is_causal, attn_mask=attn_mask
            )
        else:
            attn = (q @ k.transpose(-2, -1)) * self.scale
@@ -333,7 +326,6 @@ class ActionConditionedVideoPredictor(nn.Module):
        num_heads: int,
        mlp_ratio: float,
        num_action_tokens_per_step: int,
-        dropout: float = 0.0,
        use_extrinsics: bool = False,
    ) -> None:
        super().__init__()
@@ -341,14 +333,8 @@ class ActionConditionedVideoPredictor(nn.Module):
        self.use_extrinsics = use_extrinsics
        self.predictor_embed = nn.Linear(embed_dim, predictor_embed_dim, bias=True)
        self.action_encoder = nn.Linear(action_embed_dim, predictor_embed_dim, bias=True)
-        # `extrinsics_encoder` only feeds the `use_extrinsics` branch of `forward`. Building it
-        # unconditionally left ~2.1M parameters that never received a gradient (and were only
-        # tolerated by the trainer's global `find_unused_parameters=True`). There was a
-        # `state_encoder` here too, which nothing ever called. Checkpoints written before this
-        # change carry both; they surface as unexpected keys and are ignored on load.
-        self.extrinsics_encoder = (
-            nn.Linear(action_embed_dim - 1, predictor_embed_dim, bias=True) if use_extrinsics else None
-        )
+        self.state_encoder = nn.Linear(action_embed_dim, predictor_embed_dim, bias=True)
+        self.extrinsics_encoder = nn.Linear(action_embed_dim - 1, predictor_embed_dim, bias=True)

        self.img_height, self.img_width = img_size
        self.patch_size = patch_size
@@ -364,8 +350,8 @@ class ActionConditionedVideoPredictor(nn.Module):
                    num_heads=num_heads,
                    mlp_ratio=mlp_ratio,
                    qkv_bias=True,
-                    drop=dropout,
-                    attn_drop=dropout,
+                    drop=0.0,
+                    attn_drop=0.0,
                    drop_path=0.0,
                    norm_layer=lambda dim: nn.LayerNorm(dim, eps=1e-6),
                    grid_size=self.grid_height,
@@ -175,9 +175,6 @@ class AddBatchDimensionComplementaryDataStep(ComplementaryDataProcessorStep):
            if isinstance(task_index_value, Tensor) and task_index_value.dim() == 0:
                complementary_data["task_index"] = task_index_value.unsqueeze(0)

-        complementary_data.pop("language_persistent", None)
-        complementary_data.pop("language_events", None)
-
        if "messages" in complementary_data:
            messages = complementary_data["messages"]
            if isinstance(messages, list) and (not messages or isinstance(messages[0], dict)):
@@ -647,10 +647,15 @@ def main():
    tags = set(tags).union({"robotics", "lerobot", policy_type})
    tags = list(tags)

-    # Generate model card
-    card = policy.generate_model_card(
-        dataset_repo_id=dataset_repo_id, model_type=policy_type, license=license, tags=tags
-    )
+    # Generate model card through the free helper (PreTrainedPolicy.generate_model_card was
+    # removed with the publisher redesign), then apply the metadata recovered above — the
+    # migrated policy config does not carry the original repo's card fields.
+    from lerobot.common.train_utils import generate_model_card
+
+    card = generate_model_card(policy.config)
+    card.data.datasets = dataset_repo_id
+    card.data.license = license
+    card.data.tags = sorted(tags)

    # Save model card locally
    card.save(str(output_dir / "README.md"))
@@ -41,7 +41,7 @@ from pathlib import Path
 from typing import Any, TypedDict, TypeVar, cast

 import torch
-from huggingface_hub import hf_hub_download
+from huggingface_hub import hf_hub_download, snapshot_download
 from safetensors.torch import load_file, save_file

 from lerobot.configs import FeatureType, PipelineFeatureType, PolicyFeature
@@ -212,6 +212,10 @@ class ProcessorStep(ABC):
        """
        return None

+    def save_artifacts(self, save_directory: Path) -> dict[str, str]:
+        """Save non-tensor assets and map constructor arguments to relative paths."""
+        return {}
+
    def reset(self) -> None:
        """Resets the internal state of the processor step, if any."""
        return None
@@ -556,6 +560,22 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
        pipeline_config = self.get_config()
        pipeline_state_dict = self.state_dict()

+        for processor_step, step_entry in zip(self.steps, pipeline_config["steps"], strict=True):
+            artifacts = processor_step.save_artifacts(save_directory)
+            if artifacts:
+                for config_key, relative_path in artifacts.items():
+                    artifact_path = Path(relative_path)
+                    if artifact_path.is_absolute() or ".." in artifact_path.parts:
+                        raise ValueError(
+                            f"Processor artifact path must be relative to the checkpoint: {relative_path!r}"
+                        )
+                    if not (save_directory / artifact_path).exists():
+                        raise FileNotFoundError(
+                            f"Processor step did not save declared artifact '{relative_path}'"
+                        )
+                    step_entry["config"][config_key] = artifact_path.as_posix()
+                step_entry["artifacts"] = artifacts
+
        for state_key, step_state_dict in pipeline_state_dict.items():
            state_filename = f"{state_key}.safetensors"
            save_file(step_state_dict, save_directory / state_filename)
@@ -740,7 +760,13 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):

        # 3. Build steps with overrides
        steps, validated_overrides = cls._build_steps_with_overrides(
-            loaded_config, overrides or {}, model_id, base_path, hub_download_kwargs, is_local_source
+            loaded_config,
+            overrides or {},
+            model_id,
+            base_path,
+            config_filename,
+            hub_download_kwargs,
+            is_local_source,
        )

        # 4. Validate that all overrides were used
@@ -936,6 +962,7 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
        overrides: dict[str, Any],
        model_id: str,
        base_path: Path | None,
+        config_filename: str,
        hub_download_kwargs: dict[str, Any],
        is_local_source: bool = False,
    ) -> tuple[list[ProcessorStep], set[str]]:
@@ -945,6 +972,11 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):

        **For each step in loaded_config["steps"]**:

+        0. **Artifact Resolution** (via _resolve_artifact_paths):
+           - Resolve declared relative artifact paths against a local checkpoint
+           - Download declared artifacts when loading the pipeline from the Hub
+           - Reject absolute paths and path traversal before step construction
+
        1. **Class Resolution** (via _resolve_step_class):
           - **If "registry_name" exists**: Look up in ProcessorStepRegistry
             Example: {"registry_name": "normalize_step"} -> Get registered class
@@ -978,6 +1010,8 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
            overrides: User-provided parameter overrides (keyed by class/registry name)
            model_id: The model identifier (needed for Hub state file downloads)
            base_path: Local directory path for finding state files
+            config_filename: Processor config path, used as the repository-relative
+                base for state files and declared artifacts.
            hub_download_kwargs: Parameters for hf_hub_download (tokens, cache, etc.)
            is_local_source: Whether model_id resolved to a local directory or config file.

@@ -990,15 +1024,80 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
            ImportError: If a step class cannot be imported or found in registry
            ValueError: If a step cannot be instantiated with its configuration
        """
+        loaded_config = deepcopy(loaded_config)
+        cls._resolve_artifact_paths(
+            loaded_config,
+            model_id,
+            base_path,
+            config_filename,
+            hub_download_kwargs,
+        )
        steps, remaining_override_keys = cls._build_steps_from_config(loaded_config, overrides)

        for step_instance, step_entry in zip(steps, loaded_config["steps"], strict=True):
            cls._load_step_state(
-                step_instance, step_entry, model_id, base_path, hub_download_kwargs, is_local_source
+                step_instance,
+                step_entry,
+                model_id,
+                base_path,
+                config_filename,
+                hub_download_kwargs,
+                is_local_source,
            )

        return steps, remaining_override_keys

+    @classmethod
+    def _resolve_artifact_paths(
+        cls,
+        loaded_config: dict[str, Any],
+        model_id: str,
+        base_path: Path | None,
+        config_filename: str,
+        hub_download_kwargs: dict[str, Any],
+    ) -> None:
+        """Resolve declared relative processor artifacts before step construction.
+
+        Args:
+            loaded_config: Mutable processor configuration containing step artifact declarations.
+            model_id: Local checkpoint path or Hub model identifier.
+            base_path: Local directory containing the resolved processor configuration.
+            config_filename: Processor config path, whose parent is the artifact root on the Hub.
+            hub_download_kwargs: Authentication, revision, and cache arguments for Hub downloads.
+
+        Raises:
+            ValueError: If a declared artifact path is absolute or escapes the checkpoint.
+            FileNotFoundError: If a declared artifact cannot be found locally or downloaded.
+        """
+        is_local = Path(model_id).is_dir() or Path(model_id).is_file()
+
+        for step_entry in loaded_config["steps"]:
+            artifacts = step_entry.get("artifacts", {})
+            for config_key, relative_path in artifacts.items():
+                artifact_path = Path(relative_path)
+                if artifact_path.is_absolute() or ".." in artifact_path.parts:
+                    raise ValueError(
+                        f"Processor artifact path must be relative to the checkpoint: {relative_path!r}"
+                    )
+
+                resolved_path = base_path / artifact_path if base_path is not None else artifact_path
+                if not resolved_path.exists() and not is_local:
+                    repository_path = Path(config_filename).parent / artifact_path
+                    snapshot_download(
+                        repo_id=model_id,
+                        repo_type="model",
+                        allow_patterns=f"{repository_path.as_posix()}/**",
+                        **hub_download_kwargs,
+                    )
+
+                if not resolved_path.exists():
+                    step_name = step_entry.get("registry_name", step_entry.get("class", "unknown"))
+                    raise FileNotFoundError(
+                        f"Missing processor artifact '{relative_path}' for step '{step_name}' "
+                        f"next to '{config_filename}'. Checkpoint artifacts are incomplete."
+                    )
+                step_entry["config"][config_key] = str(resolved_path)
+
    @classmethod
    def _build_steps_from_config(
        cls,
@@ -1158,6 +1257,7 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
        step_entry: dict[str, Any],
        model_id: str,
        base_path: Path | None,
+        config_filename: str,
        hub_download_kwargs: dict[str, Any],
        is_local_source: bool = False,
    ) -> None:
@@ -1198,6 +1298,8 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
            step_entry: The step configuration dictionary (may contain "state_file")
            model_id: The model identifier (used for Hub downloads if needed)
            base_path: Local directory path for finding state files (None for Hub-only)
+            config_filename: Processor config path, whose parent is used to resolve
+                repository-relative state files on the Hub.
            hub_download_kwargs: Parameters for hf_hub_download (tokens, cache, etc.)
            is_local_source: Whether model_id resolved to a local directory or config file.

@@ -1223,7 +1325,7 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
            # Download from Hub
            state_path = hf_hub_download(
                repo_id=model_id,
-                filename=state_filename,
+                filename=(Path(config_filename).parent / state_filename).as_posix(),
                repo_type="model",
                **hub_download_kwargs,
            )
@@ -51,12 +51,6 @@ def to_relative_actions(actions: Tensor, state: Tensor, mask: Sequence[bool]) ->
    # DeviceProcessorStep moves the transition, so it can be on CPU while actions are on CUDA.
    if state.device != actions.device or state.dtype != actions.dtype:
        state = state.to(device=actions.device, dtype=actions.dtype)
-    # When the observation is temporally stacked (e.g. LingBot-VA loads several obs steps via
-    # observation_delta_indices, giving state shape (B, T_obs, state_dim)), the relative reference
-    # is the CURRENT frame (delta == 0, i.e. index 0). Collapse to it so the offset broadcasts over
-    # the action horizon. pi0/pi05 pass a 2D (B, state_dim) state and are unaffected.
-    if state.ndim == 3:
-        state = state[:, 0]
    state_offset = state[..., :dims] * mask_t
    if actions.ndim == 3:
        state_offset = state_offset.unsqueeze(-2)
@@ -79,10 +73,6 @@ def to_absolute_actions(actions: Tensor, state: Tensor, mask: Sequence[bool]) ->
    # DeviceProcessorStep moves the transition, so it can be on CPU while actions are on CUDA.
    if state.device != actions.device or state.dtype != actions.dtype:
        state = state.to(device=actions.device, dtype=actions.dtype)
-    # Mirror to_relative_actions: collapse a temporally-stacked (B, T_obs, state_dim) state to the
-    # current frame (index 0) so the round-trip stays symmetric with the relative conversion.
-    if state.ndim == 3:
-        state = state[:, 0]
    state_offset = state[..., :dims] * mask_t
    if actions.ndim == 3:
        state_offset = state_offset.unsqueeze(-2)
@@ -136,7 +126,7 @@ class RelativeActionsProcessorStep(ProcessorStep):
        observation = transition.get(TransitionKey.OBSERVATION, {})
        state = observation.get(OBS_STATE) if observation else None

-        # Always cache state for the paired AbsoluteActionsProcessorStep.
+        # Always cache state for the paired AbsoluteActionsProcessorStep
        if state is not None:
            self._last_state = state

@@ -156,11 +146,6 @@ class RelativeActionsProcessorStep(ProcessorStep):
        """Return the cached ``observation.state`` used as the reference point for relative/absolute action conversions."""
        return self._last_state

-    def set_cached_state(self, state: torch.Tensor | None) -> None:
-        """Override the cached anchor state, e.g. to re-pin a chunk's anchor after the
-        per-tick pipeline overwrote it (see ``SyncInferenceEngine``)."""
-        self._last_state = state
-
    def get_config(self) -> dict[str, Any]:
        return {
            "enabled": self.enabled,
@@ -16,9 +16,11 @@

 from __future__ import annotations

-from dataclasses import dataclass
+from dataclasses import asdict, dataclass
 from typing import Any

+import numpy as np
+
 from lerobot.configs import PipelineFeatureType, PolicyFeature
 from lerobot.configs.recipe import TrainingRecipe
 from lerobot.datasets.language import LANGUAGE_EVENTS, LANGUAGE_PERSISTENT
@@ -32,25 +34,46 @@ from .pipeline import ProcessorStep, ProcessorStepRegistry
@dataclass
@ProcessorStepRegistry.register(name="render_messages_processor")
 class RenderMessagesStep(ProcessorStep):
-    """Processor step that turns raw language columns into rendered chat messages.
+    """Turn raw language columns into recipe-defined messages and supervision.

-    Reads ``language_persistent`` and ``language_events`` from the transition's
-    complementary data, renders them through ``recipe`` at the sample timestamp,
-    and replaces the raw columns with the resulting ``messages`` /
-    ``message_streams`` / ``target_message_indices`` keys.
+    Reads ``language_persistent`` and ``language_events`` from complementary
+    data, renders them at each sample timestamp, and replaces the raw columns
+    with ``messages``, ``message_streams``, and ``target_message_indices``.
+    Batched inputs are filtered to samples with applicable supervision; samples
+    without language annotations use their task string as low-level supervision
+    when one is available.
    """

    recipe: TrainingRecipe
    dataset_ctx: Any | None = None

+    def __post_init__(self) -> None:
+        if isinstance(self.recipe, dict):
+            self.recipe = TrainingRecipe.from_dict(self.recipe)
+
+    def get_config(self) -> dict[str, Any]:
+        return {"recipe": asdict(self.recipe)}
+
    def __call__(self, transition: EnvTransition) -> EnvTransition | None:
-        """Render messages for a single transition; return ``None`` to drop it."""
+        """Render messages, preserving unannotated samples and dropping unmatched annotated ones."""
        complementary_data = transition.get(TransitionKey.COMPLEMENTARY_DATA) or {}
        persistent = complementary_data.get(LANGUAGE_PERSISTENT) or []
        events = complementary_data.get(LANGUAGE_EVENTS) or []

        if not persistent and not events:
-            return transition
+            # A dataset without language annotations remains usable: render its
+            # task as low-level supervision, or pass it through when no task exists.
+            rendered = _fallback_low_level_render(complementary_data.get("task"))
+            if rendered is None:
+                return transition
+            new_transition = transition.copy()
+            new_complementary_data = dict(new_transition.get(TransitionKey.COMPLEMENTARY_DATA) or {})
+            new_complementary_data.update(rendered)
+            new_transition[TransitionKey.COMPLEMENTARY_DATA] = new_complementary_data
+            return new_transition
+
+        if _is_batched_language(persistent) or _is_batched_language(events):
+            return self._call_batch(transition, complementary_data, persistent, events)

        timestamp = complementary_data.get("timestamp")
        if timestamp is None:
@@ -67,18 +90,171 @@ class RenderMessagesStep(ProcessorStep):
            dataset_ctx=self.dataset_ctx,
        )
        if rendered is None:
-            return None
+            # Language is present but this sparse frame has no applicable recipe
+            # branch. Keep it only when task-level action supervision is possible.
+            rendered = _fallback_low_level_render(complementary_data.get("task"))
+            if rendered is None:
+                return None

        new_transition = transition.copy()
-        new_complementary_data = dict(complementary_data)
+        new_complementary_data = dict(new_transition.get(TransitionKey.COMPLEMENTARY_DATA) or {})
        new_complementary_data.pop(LANGUAGE_PERSISTENT, None)
        new_complementary_data.pop(LANGUAGE_EVENTS, None)
        new_complementary_data.update(rendered)
        new_transition[TransitionKey.COMPLEMENTARY_DATA] = new_complementary_data
        return new_transition

+    def _call_batch(
+        self,
+        transition: EnvTransition,
+        complementary_data: dict[str, Any],
+        persistent_batch: list,
+        events_batch: list,
+    ) -> EnvTransition | None:
+        """Render a language batch.
+
+        Non-empty persistent and event batches must have the same size. Either
+        list may be empty when that language column is absent from the batch.
+        """
+        timestamp = complementary_data.get("timestamp")
+        if timestamp is None:
+            raise KeyError("RenderMessagesStep requires sample timestamp in complementary data.")
+
+        non_empty_batch_sizes = {len(batch) for batch in (persistent_batch, events_batch) if batch}
+        if len(non_empty_batch_sizes) > 1:
+            raise ValueError(
+                "Batched language columns must have equal lengths when both are non-empty, "
+                f"got persistent={len(persistent_batch)} and events={len(events_batch)}."
+            )
+        batch_size = next(iter(non_empty_batch_sizes), 0)
+        messages: list[list[dict[str, Any]]] = []
+        message_streams: list[list[str | None]] = []
+        target_message_indices: list[list[int]] = []
+        keep_indices: list[int] = []
+
+        for i in range(batch_size):
+            rendered = render_sample(
+                recipe=self.recipe,
+                persistent=persistent_batch[i] if i < len(persistent_batch) else [],
+                events=events_batch[i] if i < len(events_batch) else [],
+                t=_batch_value(timestamp, i),
+                sample_idx=int(_batch_value(complementary_data.get("index", 0), i)),
+                task=_batch_value(complementary_data.get("task"), i),
+                dataset_ctx=self.dataset_ctx,
+            )
+            if rendered is None:
+                rendered = _fallback_low_level_render(_batch_value(complementary_data.get("task"), i))
+                if rendered is None:
+                    continue
+            keep_indices.append(i)
+            messages.append(rendered["messages"])
+            message_streams.append(rendered["message_streams"])
+            target_message_indices.append(rendered["target_message_indices"])
+
+        if not messages:
+            return None
+
+        new_transition = (
+            _select_batch_indices(transition, keep_indices, batch_size)
+            if len(keep_indices) != batch_size
+            else transition.copy()
+        )
+        new_complementary_data = dict(new_transition.get(TransitionKey.COMPLEMENTARY_DATA) or {})
+        new_complementary_data.pop(LANGUAGE_PERSISTENT, None)
+        new_complementary_data.pop(LANGUAGE_EVENTS, None)
+        new_complementary_data["messages"] = messages
+        new_complementary_data["message_streams"] = message_streams
+        new_complementary_data["target_message_indices"] = target_message_indices
+        new_transition[TransitionKey.COMPLEMENTARY_DATA] = new_complementary_data
+        return new_transition
+
    def transform_features(
        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
        """Pass features through unchanged; rendering only touches complementary data."""
        return features
+
+
+def _is_batched_language(value: Any) -> bool:
+    return isinstance(value, list) and bool(value) and isinstance(value[0], list)
+
+
+def _batch_value(value: Any, index: int) -> Any:
+    if value is None:
+        return None
+    if isinstance(value, list):
+        return value[index]
+    if hasattr(value, "ndim") and value.ndim > 0:
+        return unwrap_scalar(value[index])
+    return unwrap_scalar(value)
+
+
+def _select_batch_indices(transition: EnvTransition, indices: list[int], batch_size: int) -> EnvTransition:
+    selected = transition.copy()
+    for key in (TransitionKey.OBSERVATION, TransitionKey.COMPLEMENTARY_DATA):
+        data = selected.get(key)
+        if isinstance(data, dict):
+            selected[key] = {
+                name: _select_value(value, indices, batch_size, f"{key}.{name}")
+                for name, value in data.items()
+            }
+    action = selected.get(TransitionKey.ACTION)
+    if action is not None:
+        selected[TransitionKey.ACTION] = _select_value(action, indices, batch_size, str(TransitionKey.ACTION))
+    return selected
+
+
+def _select_value(value: Any, indices: list[int], batch_size: int, path: str) -> Any:
+    if isinstance(value, dict):
+        return {key: _select_value(item, indices, batch_size, f"{path}.{key}") for key, item in value.items()}
+    if isinstance(value, list):
+        if len(value) != batch_size:
+            raise ValueError(
+                f"Cannot filter batched field {path!r}: expected {batch_size} values, got {len(value)}."
+            )
+        return [value[i] for i in indices]
+    if isinstance(value, np.ndarray) and value.ndim > 0:
+        return value[indices]
+    if hasattr(value, "index_select") and hasattr(value, "new_tensor") and getattr(value, "ndim", 0) > 0:
+        return value.index_select(0, value.new_tensor(indices).long())
+    return value
+
+
+def _fallback_low_level_render(task: Any) -> dict[str, Any] | None:
+    """Keep action-only samples trainable when no recipe branch matches."""
+    if hasattr(task, "item"):
+        task = task.item()
+    if isinstance(task, list):
+        if not task:
+            return None
+        messages = []
+        message_streams = []
+        target_message_indices = []
+        missing_indices = []
+        for index, t in enumerate(task):
+            rendered = _fallback_low_level_render(t)
+            if rendered is None:
+                missing_indices.append(index)
+                continue
+            messages.append(rendered["messages"])
+            message_streams.append(rendered["message_streams"])
+            target_message_indices.append(rendered["target_message_indices"])
+        if missing_indices:
+            if len(missing_indices) == len(task):
+                return None
+            raise ValueError(
+                "Batched low-level fallback requires a non-empty task for every sample; "
+                f"missing task at indices {missing_indices}."
+            )
+        return {
+            "messages": messages,
+            "message_streams": message_streams,
+            "target_message_indices": target_message_indices,
+        }
+    if not isinstance(task, str) or not task:
+        return None
+    return {
+        "messages": [{"role": "user", "content": task}],
+        "message_streams": ["low_level"],
+        "target_message_indices": [],
+    }
@@ -25,6 +25,7 @@ from __future__ import annotations

 import logging
 from dataclasses import dataclass, field
+from pathlib import Path
 from typing import TYPE_CHECKING, Any

 import torch
@@ -32,6 +33,7 @@ import torch
 from lerobot.configs import FeatureType, PipelineFeatureType, PolicyFeature
 from lerobot.lerobot_types import EnvTransition, RobotObservation, TransitionKey
 from lerobot.utils.constants import (
+    ACTION_CODE_TOKEN_MASK,
    ACTION_TOKEN_MASK,
    ACTION_TOKENS,
    OBS_LANGUAGE_ATTENTION_MASK,
@@ -136,7 +138,7 @@ class TokenizerProcessorStep(ObservationProcessorStep):
        # Standardize to a list of strings for the tokenizer
        if isinstance(task, str):
            return [task]
-        elif isinstance(task, (list, tuple)) and all(isinstance(t, str) for t in task):
+        elif isinstance(task, list | tuple) and all(isinstance(t, str) for t in task):
            return list(task)

        return None
@@ -293,6 +295,15 @@ class TokenizerProcessorStep(ObservationProcessorStep):

        return config

+    def save_artifacts(self, save_directory: Path) -> dict[str, str]:
+        """Save the tokenizer so object-provided instances reload without overrides."""
+        artifact_path = Path("tokenizer")
+        save_pretrained = getattr(self.input_tokenizer, "save_pretrained", None)
+        if save_pretrained is None:
+            raise TypeError("Tokenizer must implement save_pretrained() to save a portable pipeline.")
+        save_pretrained(save_directory / artifact_path)
+        return {"tokenizer_name": artifact_path.as_posix()}
+
    def transform_features(
        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
@@ -349,6 +360,7 @@ class ActionTokenizerProcessorStep(ActionProcessorStep):
    max_action_tokens: int = 256
    fast_skip_tokens: int = 128
    paligemma_tokenizer_name: str = "google/paligemma-3b-pt-224"
+    allow_truncation: bool = True
    # Internal tokenizer instance (not part of the config)
    action_tokenizer: Any = field(default=None, init=False, repr=False)
    _paligemma_tokenizer: Any = field(default=None, init=False, repr=False)
@@ -412,14 +424,15 @@ class ActionTokenizerProcessorStep(ActionProcessorStep):
            # During inference, no action is available, skip tokenization
            return new_transition

-        # Tokenize and get both tokens and mask
-        tokens, mask = self._tokenize_action(action)
+        # Tokenize and get masks for the full formatted sequence and the discrete action codes.
+        tokens, mask, code_mask = self._tokenize_action(action)

        # Store mask in complementary data
        complementary_data = new_transition.get(TransitionKey.COMPLEMENTARY_DATA, {})
        if complementary_data is None:
            complementary_data = {}
        complementary_data[ACTION_TOKEN_MASK] = mask
+        complementary_data[ACTION_CODE_TOKEN_MASK] = code_mask
        complementary_data[ACTION_TOKENS] = tokens
        new_transition[TransitionKey.COMPLEMENTARY_DATA] = complementary_data
        return new_transition
@@ -430,7 +443,7 @@ class ActionTokenizerProcessorStep(ActionProcessorStep):
        """
        return self._paligemma_tokenizer.vocab_size - 1 - self.fast_skip_tokens - tokens

-    def _tokenize_action(self, action: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+    def _tokenize_action(self, action: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        """
        Tokenizes the action tensor and creates a mask.

@@ -459,6 +472,7 @@ class ActionTokenizerProcessorStep(ActionProcessorStep):
        # The fast tokenizer expects action data and returns token IDs
        tokens_list = []
        masks_list = []
+        code_masks_list = []

        for i in range(batch_size):
            # Tokenize single action (move to CPU first as tokenizer uses scipy which requires numpy)
@@ -476,65 +490,82 @@ class ActionTokenizerProcessorStep(ActionProcessorStep):
            if tokens.dim() > 1:
                tokens = tokens.flatten()

+            action_code_tokens = self._act_tokens_to_paligemma_tokens(tokens)
            bos_id = self._paligemma_tokenizer.bos_token_id
-            # add bos
+            prompt_tokens = torch.tensor(
+                self._paligemma_tokenizer.encode("Action: ", add_special_tokens=False),
+                device=action.device,
+            )
+            end_tokens = torch.tensor(self._paligemma_tokenizer.encode("|"), device=action.device)
+
+            code_start = 1 + len(prompt_tokens)
+            code_end = code_start + len(action_code_tokens)
            tokens = torch.cat(
                [
                    torch.tensor([bos_id], device=action.device),
-                    torch.tensor(
-                        self._paligemma_tokenizer.encode("Action: ", add_special_tokens=False),
-                        device=action.device,
-                    ),
-                    self._act_tokens_to_paligemma_tokens(tokens),
-                    torch.tensor(self._paligemma_tokenizer.encode("|"), device=action.device),
+                    prompt_tokens,
+                    action_code_tokens,
+                    end_tokens,
                ]
            )
+            code_mask = torch.zeros(len(tokens), dtype=torch.bool, device=action.device)
+            code_mask[code_start:code_end] = True

            # Truncate or pad to max_action_tokens
            if len(tokens) > self.max_action_tokens:
+                if not self.allow_truncation:
+                    raise ValueError(
+                        f"FAST action sequence has {len(tokens)} tokens, exceeding "
+                        f"max_action_tokens={self.max_action_tokens}."
+                    )
                logging.warning(
                    f"Token length ({len(tokens)}) exceeds max length ({self.max_action_tokens}), truncating. "
                    "Consider increasing the `max_action_tokens` in your model config if this happens frequently."
                )
                tokens = tokens[: self.max_action_tokens]
+                code_mask = code_mask[: self.max_action_tokens]
                mask = torch.ones(self.max_action_tokens, dtype=torch.bool, device=action.device)
            else:
+                pad_len = self.max_action_tokens - len(tokens)
                mask = torch.cat(
                    [
                        torch.ones(len(tokens), dtype=torch.bool, device=action.device),
-                        torch.zeros(
-                            self.max_action_tokens - len(tokens), dtype=torch.bool, device=action.device
-                        ),
+                        torch.zeros(pad_len, dtype=torch.bool, device=action.device),
                    ]
                )
+                code_mask = torch.nn.functional.pad(code_mask, (0, pad_len), value=False)
                # Pad tokens with zeros
-                tokens = torch.nn.functional.pad(tokens, (0, self.max_action_tokens - len(tokens)), value=0)
+                tokens = torch.nn.functional.pad(tokens, (0, pad_len), value=0)

            tokens_list.append(tokens)
            masks_list.append(mask)
+            code_masks_list.append(code_mask)

        # Stack into batched tensors
        tokens_batch = torch.stack(tokens_list, dim=0)  # (B, max_action_tokens)
        masks_batch = torch.stack(masks_list, dim=0)  # (B, max_action_tokens)
+        code_masks_batch = torch.stack(code_masks_list, dim=0)  # (B, max_action_tokens)

        # Remove batch dimension if input was single sample
        if single_sample:
            tokens_batch = tokens_batch.squeeze(0)
            masks_batch = masks_batch.squeeze(0)
+            code_masks_batch = code_masks_batch.squeeze(0)

        # Move to the same device as the input
        if device is not None:
            tokens_batch = tokens_batch.to(device)
            masks_batch = masks_batch.to(device)
+            code_masks_batch = code_masks_batch.to(device)

-        return tokens_batch, masks_batch
+        return tokens_batch, masks_batch, code_masks_batch

    def action(self, action: torch.Tensor) -> torch.Tensor:
        """
        This method is not used since we override __call__.
        Required by ActionProcessorStep ABC.
        """
-        tokens, _ = self._tokenize_action(action)
+        tokens, _, _ = self._tokenize_action(action)
        return tokens

    def get_config(self) -> dict[str, Any]:
@@ -550,6 +581,9 @@ class ActionTokenizerProcessorStep(ActionProcessorStep):
        config = {
            "trust_remote_code": self.trust_remote_code,
            "max_action_tokens": self.max_action_tokens,
+            "fast_skip_tokens": self.fast_skip_tokens,
+            "paligemma_tokenizer_name": self.paligemma_tokenizer_name,
+            "allow_truncation": self.allow_truncation,
        }

        # Only save tokenizer_name if it was used to create the tokenizer
@@ -558,6 +592,14 @@ class ActionTokenizerProcessorStep(ActionProcessorStep):

        return config

+    def save_artifacts(self, save_directory: Path) -> dict[str, str]:
+        artifact_path = Path("action_tokenizer")
+        save_pretrained = getattr(self.action_tokenizer, "save_pretrained", None)
+        if save_pretrained is None:
+            raise TypeError("Action tokenizer must implement save_pretrained() to save a portable pipeline.")
+        save_pretrained(save_directory / artifact_path)
+        return {"action_tokenizer_name": artifact_path.as_posix()}
+
    def transform_features(
        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
@@ -16,12 +16,11 @@ import abc
 import builtins
 import logging
 import os
-from importlib.resources import files
+import warnings
 from pathlib import Path
-from tempfile import TemporaryDirectory
 from typing import TYPE_CHECKING, Any, TypeVar

-from huggingface_hub import HfApi, ModelCard, ModelCardData, hf_hub_download
+from huggingface_hub import hf_hub_download
 from huggingface_hub.constants import SAFETENSORS_SINGLE_FILE
 from huggingface_hub.errors import HfHubHTTPError
 from safetensors.torch import load_model as load_model_as_safetensor, save_model as save_model_as_safetensor
@@ -61,6 +60,22 @@ class PreTrainedRewardModel(nn.Module, HubMixin, abc.ABC):
            raise TypeError(f"Class {cls.__name__} must define 'name'")

    def _save_pretrained(self, save_directory: Path) -> None:
+        """Serialize this reward model's parameters (and config) into `save_directory`.
+
+        Safe to call on every rank: replicas carry identical weights, so only the main process
+        writes (sharded reward models are rejected at config validation — no collective gather).
+
+        Args:
+            save_directory (Path): Target directory for the reward model config (`config.json`)
+                and `model.safetensors`.
+        """
+        from lerobot.distributed.utils import is_main_process
+
+        # save_checkpoint calls this on every rank; replicas carry identical
+        # weights, so the main process is the only writer. Sharded reward models are rejected
+        # at config validation, so no collective gather is needed here.
+        if not is_main_process():
+            return
        self.config._save_pretrained(save_directory)
        model_to_save = self.module if hasattr(self, "module") else self
        save_model_as_safetensor(model_to_save, str(save_directory / SAFETENSORS_SINGLE_FILE))
@@ -175,53 +190,22 @@ class PreTrainedRewardModel(nn.Module, HubMixin, abc.ABC):
        """
        return type(self).forward is not PreTrainedRewardModel.forward

-    def push_model_to_hub(self, cfg: "TrainPipelineConfig"):
-        api = HfApi()
-        repo_id = api.create_repo(
-            repo_id=self.config.repo_id, private=self.config.private, exist_ok=True
-        ).repo_id
+    def push_model_to_hub(self, cfg: "TrainPipelineConfig") -> None:
+        """Publish this reward model to the Hub.

-        # Push the files to the repo in a single commit
-        with TemporaryDirectory(ignore_cleanup_errors=True) as tmp:
-            saved_path = Path(tmp) / repo_id
+        Deprecated: use :func:`lerobot.common.train_utils.publish_trained_model` instead.

-            self.save_pretrained(saved_path)  # Calls _save_pretrained and stores model tensors
+        Args:
+            cfg (TrainPipelineConfig): The training config; saved as `train_config.json` and
+                used to render the model card.
+        """
+        from lerobot.common.train_utils import publish_trained_model

-            card = self.generate_model_card(
-                cfg.dataset.repo_id, self.config.type, self.config.license, self.config.tags
-            )
-            card.save(str(saved_path / "README.md"))
-
-            cfg.save_pretrained(saved_path)  # Calls _save_pretrained and stores train config
-
-            commit_info = api.upload_folder(
-                repo_id=repo_id,
-                repo_type="model",
-                folder_path=saved_path,
-                commit_message="Upload reward model weights, train config and readme",
-                allow_patterns=["*.safetensors", "*.json", "*.yaml", "*.md"],
-                ignore_patterns=["*.tmp", "*.log"],
-            )
-
-            logging.info(f"Model pushed to {commit_info.repo_url.url}")
-
-    def generate_model_card(
-        self, dataset_repo_id: str, model_type: str, license: str | None, tags: list[str] | None
-    ) -> ModelCard:
-        card_data = ModelCardData(
-            license=license or "apache-2.0",
-            library_name="lerobot",
-            pipeline_tag="robotics",
-            tags=list(set(tags or []).union({"robotics", "lerobot", "reward-model", model_type})),
-            model_name=model_type,
-            datasets=dataset_repo_id,
+        warnings.warn(
+            "PreTrainedRewardModel.push_model_to_hub is deprecated and will be removed in a "
+            "future version. Use lerobot.common.train_utils.publish_trained_model(cfg, model, "
+            "preprocessor, postprocessor, dataset_meta) instead.",
+            FutureWarning,
+            stacklevel=2,
        )
-
-        template_card = (
-            files("lerobot.templates")
-            .joinpath("lerobot_rewardmodel_modelcard_template.md")
-            .read_text(encoding="utf-8")
-        )
-        card = ModelCard.from_template(card_data, template_str=template_card)
-        card.validate()
-        return card
+        publish_trained_model(cfg, self, None, None, None)
@@ -16,6 +16,7 @@

 from __future__ import annotations

+import logging
 import random
 from typing import TYPE_CHECKING, Any

@@ -69,6 +70,8 @@ from .sarm_utils import (
    pad_state_to_max_dim,
 )

+logger = logging.getLogger(__name__)
+

 class SARMEncodingProcessorStep(ProcessorStep):
    """ProcessorStep that encodes images and text with CLIP and generates stage and progress labels for SARM."""
@@ -108,6 +111,8 @@ class SARMEncodingProcessorStep(ProcessorStep):
            else None
        )

+        self._validate_annotation_columns()
+
        self.device = torch.device(
            self.config.device if self.config.device else "cuda" if torch.cuda.is_available() else "cpu"
        )
@@ -120,6 +125,78 @@ class SARMEncodingProcessorStep(ProcessorStep):
        self.verbs = ["move", "grasp", "rotate", "push", "pull", "slide", "lift", "place"]
        self.fake = Faker()

+    @staticmethod
+    def _resolve_annotation_column(episodes_df: pd.DataFrame, annotation_type: str, suffix: str) -> str:
+        """Resolve a mode-specific annotation column, falling back to the legacy unprefixed name."""
+        prefixed = f"{annotation_type}_{suffix}"
+        return prefixed if prefixed in episodes_df.columns else suffix
+
+    @staticmethod
+    def _annotations_are_usable(names: Any, starts: Any, ends: Any) -> bool:
+        """Return whether an episode has non-empty, aligned annotation arrays."""
+        values = (names, starts, ends)
+        if not all(isinstance(value, (list, tuple, np.ndarray)) for value in values):
+            return False
+
+        lengths = {len(value) for value in values}
+        return len(lengths) == 1 and next(iter(lengths)) > 0
+
+    def _validate_annotation_columns(self) -> None:
+        """Validate annotation coverage before loading models or generating training targets.
+
+        A multi-stage head with no usable episode annotations would otherwise train entirely
+        against all-zero targets. Reject that configuration and warn when only part of the
+        dataset is usable.
+        """
+        if self.dataset_meta is None:
+            return
+        episodes_df = self.dataset_meta.episodes.to_pandas()
+        num_episodes = len(episodes_df)
+
+        modes = []
+        if self.dense_subtask_names and len(self.dense_subtask_names) > 1:
+            modes.append(("dense", self.dense_subtask_names))
+        if self.sparse_subtask_names and len(self.sparse_subtask_names) > 1:
+            modes.append(("sparse", self.sparse_subtask_names))
+
+        for annotation_type, names in modes:
+            columns = [
+                self._resolve_annotation_column(episodes_df, annotation_type, suffix)
+                for suffix in ("subtask_names", "subtask_start_frames", "subtask_end_frames")
+            ]
+            missing_columns = [column for column in columns if column not in episodes_df.columns]
+            if missing_columns:
+                num_usable = 0
+            else:
+                num_usable = sum(
+                    self._annotations_are_usable(*(episodes_df.loc[ep_idx, column] for column in columns))
+                    for ep_idx in episodes_df.index
+                )
+
+            if num_usable == 0:
+                missing_columns_message = (
+                    f" Missing required columns: {', '.join(missing_columns)}." if missing_columns else ""
+                )
+                raise ValueError(
+                    f"SARM {annotation_type} head is configured with {len(names)} stages, but none of "
+                    f"the {num_episodes} episodes have usable annotations in meta/episodes/*.parquet. "
+                    f"Required columns: {', '.join(columns)}.{missing_columns_message} "
+                    "Training would produce all-zero "
+                    "targets. Materialize the annotations into the episodes metadata before training."
+                )
+
+            num_unusable = num_episodes - num_usable
+            if num_unusable:
+                logger.warning(
+                    "SARM %s head: %d/%d episodes have unusable annotations in columns %s; "
+                    "their targets will be 0 and only the %d annotated episodes will train the head.",
+                    annotation_type,
+                    num_unusable,
+                    num_episodes,
+                    ", ".join(columns),
+                    num_usable,
+                )
+
    def _find_episode_for_frame(self, frame_idx: int) -> int:
        """Find the episode index for a given frame index."""
        for ep_idx in range(len(self.dataset_meta.episodes)):
@@ -167,24 +244,18 @@ class SARMEncodingProcessorStep(ProcessorStep):
        if episodes_df is None or len(global_names) == 1:
            return None, None, None

-        # Resolve column name with fallback
-        def col(suffix):
-            prefixed = f"{annotation_type}_{suffix}"
-            return prefixed if prefixed in episodes_df.columns else suffix
-
-        col_names = col("subtask_names")
-        if col_names not in episodes_df.columns or ep_idx >= len(episodes_df):
+        columns = [
+            self._resolve_annotation_column(episodes_df, annotation_type, suffix)
+            for suffix in ("subtask_names", "subtask_start_frames", "subtask_end_frames")
+        ]
+        if any(column not in episodes_df.columns for column in columns) or ep_idx >= len(episodes_df):
            return None, None, None

-        subtask_names = episodes_df.loc[ep_idx, col_names]
-        if subtask_names is None or (isinstance(subtask_names, float) and pd.isna(subtask_names)):
+        annotations = tuple(episodes_df.loc[ep_idx, column] for column in columns)
+        if not self._annotations_are_usable(*annotations):
            return None, None, None

-        return (
-            subtask_names,
-            episodes_df.loc[ep_idx, col("subtask_start_frames")],
-            episodes_df.loc[ep_idx, col("subtask_end_frames")],
-        )
+        return annotations

    def __call__(self, transition: EnvTransition) -> EnvTransition:
        """
@@ -58,12 +58,11 @@ import builtins
 import logging
 import os
 from pathlib import Path
-from tempfile import TemporaryDirectory
 from typing import TYPE_CHECKING, Any, TypeVar

 import numpy as np
 import torch
-from huggingface_hub import HfApi, hf_hub_download
+from huggingface_hub import hf_hub_download
 from huggingface_hub.constants import CONFIG_NAME
 from huggingface_hub.errors import HfHubHTTPError
 from torch import Tensor
@@ -75,9 +74,6 @@ from lerobot.rewards.topreward.configuration_topreward import TOPRewardConfig
 from lerobot.rewards.topreward.processor_topreward import TOPREWARD_FEATURE_PREFIX, TOPREWARD_INPUT_KEYS
 from lerobot.utils.import_utils import _transformers_available, require_package

-if TYPE_CHECKING:
-    from lerobot.configs.train import TrainPipelineConfig
-
 if TYPE_CHECKING or _transformers_available:
    from transformers import Qwen3VLForConditionalGeneration
 else:
@@ -205,34 +201,3 @@ class TOPRewardModel(PreTrainedRewardModel):
        instance.to(config.device)
        instance.eval()
        return instance
-
-    def push_model_to_hub(self, cfg: TrainPipelineConfig):
-        """Push the TOPReward ``config.json`` + model card to the Hub."""
-        api = HfApi()
-        repo_id = api.create_repo(
-            repo_id=self.config.repo_id, private=self.config.private, exist_ok=True
-        ).repo_id
-
-        with TemporaryDirectory(ignore_cleanup_errors=True) as tmp:
-            saved_path = Path(tmp) / repo_id
-            saved_path.mkdir(parents=True, exist_ok=True)
-
-            self.config._save_pretrained(saved_path)
-
-            card = self.generate_model_card(
-                cfg.dataset.repo_id, self.config.type, self.config.license, self.config.tags
-            )
-            card.save(str(saved_path / "README.md"))
-
-            cfg.save_pretrained(saved_path)
-
-            commit_info = api.upload_folder(
-                repo_id=repo_id,
-                repo_type="model",
-                folder_path=saved_path,
-                commit_message="Upload TOPReward config and readme",
-                allow_patterns=["*.json", "*.yaml", "*.md"],
-                ignore_patterns=["*.tmp", "*.log", "*.safetensors"],
-            )
-
-            logger.info(f"Model pushed to {commit_info.repo_url.url}")
@@ -74,13 +74,14 @@ from torch.optim.optimizer import Optimizer
 from lerobot.cameras import opencv  # noqa: F401
 from lerobot.common.train_utils import (
    get_step_checkpoint_dir,
-    load_training_state as utils_load_training_state,
+    load_training_metadata,
    save_checkpoint,
    update_last_checkpoint,
 )
 from lerobot.common.wandb_utils import WandBLogger
 from lerobot.configs import parser
 from lerobot.datasets import LeRobotDataset, make_dataset
+from lerobot.optim import load_optimizer_state
 from lerobot.policies import make_policy, make_pre_post_processors
 from lerobot.robots import so_follower  # noqa: F401
 from lerobot.teleoperators import gamepad, so_leader  # noqa: F401
@@ -103,7 +104,7 @@ from lerobot.utils.constants import (
 from lerobot.utils.device_utils import get_safe_torch_device
 from lerobot.utils.io_utils import load_json, write_json
 from lerobot.utils.process import ProcessSignalHandler, ensure_multiprocessing_start_method
-from lerobot.utils.random_utils import set_seed
+from lerobot.utils.random_utils import load_rng_state, set_seed
 from lerobot.utils.utils import (
    format_big_number,
    init_logging,
@@ -716,15 +717,18 @@ def load_training_state(
    algorithm-owned tensors) from the most recent checkpoint.

    Args:
-        cfg: Training configuration.
-        optimizers: Optimizers to load state into.
-        algorithm: Algorithm whose state dict should be restored.
-            Required for full main-equivalent resume;
-            the policy itself is restored separately via ``make_policy``.
-        device: Device on which to place loaded algorithm tensors.
+        cfg (TrainRLServerPipelineConfig): Training configuration; `cfg.resume` gates the load and
+            `cfg.output_dir` locates the last checkpoint.
+        optimizers (Optimizer | dict[str, Optimizer]): Optimizers to load state into.
+        algorithm (RLAlgorithm | None, optional): Algorithm whose state dict should be restored.
+            Required for full main-equivalent resume; the policy itself is restored separately via
+            `make_policy`. Defaults to None.
+        device (str | torch.device, optional): Device on which to place loaded algorithm tensors.
+            Defaults to "cpu".

    Returns:
-        tuple: (optimization_step, interaction_step) or (None, None) if not resuming
+        tuple[int | None, int | None]: `(optimization_step, interaction_step)`, or `(None, None)`
+        when not resuming or when loading the training state fails.
    """
    if not cfg.resume:
        return None, None
@@ -736,7 +740,10 @@ def load_training_state(

    try:
        # Restore optimizers + RNG + step from the standard `training_state/` folder
-        step, optimizers, _ = utils_load_training_state(checkpoint_dir, optimizers, None)
+        training_state_dir = checkpoint_dir / TRAINING_STATE_DIR
+        load_rng_state(training_state_dir)
+        step = load_training_metadata(training_state_dir)["step"]
+        optimizers = load_optimizer_state(optimizers, training_state_dir)

        # Restore algorithm-owned tensors
        if algorithm is not None:
@@ -58,6 +58,12 @@ class LeKiwiConfig(RobotConfig):
    # Set to `True` for backward compatibility with previous policies/dataset
    use_degrees: bool = True

+    # Number of extra attempts when a `sync_read` of the motors fails. Feetech buses can occasionally
+    # return a corrupted status packet ("Incorrect status packet!"), especially when several joints move
+    # at once, which otherwise aborts the control loop. Retries are immediate (no sleep) and only happen on
+    # failure, so the steady-state read cost is unchanged.
+    num_read_retries: int = 2
+

@dataclass
 class LeKiwiHostConfig:
@@ -347,8 +347,12 @@ class LeKiwi(Robot):
    def get_observation(self) -> RobotObservation:
        # Read actuators position for arm and vel for base
        start = time.perf_counter()
-        arm_pos = self.bus.sync_read("Present_Position", self.arm_motors)
-        base_wheel_vel = self.bus.sync_read("Present_Velocity", self.base_motors)
+        arm_pos = self.bus.sync_read(
+            "Present_Position", self.arm_motors, num_retry=self.config.num_read_retries
+        )
+        base_wheel_vel = self.bus.sync_read(
+            "Present_Velocity", self.base_motors, num_retry=self.config.num_read_retries
+        )

        base_vel = self._wheel_raw_to_body(
            base_wheel_vel["base_left_wheel"],
@@ -397,8 +401,13 @@ class LeKiwi(Robot):
        # Cap goal position when too far away from present position.
        # /!\ Slower fps expected due to reading from the follower.
        if self.config.max_relative_target is not None:
-            present_pos = self.bus.sync_read("Present_Position", self.arm_motors)
-            goal_present_pos = {key: (g_pos, present_pos[key]) for key, g_pos in arm_goal_pos.items()}
+            present_pos = self.bus.sync_read(
+                "Present_Position", self.arm_motors, num_retry=self.config.num_read_retries
+            )
+            # `arm_goal_pos` is keyed with the ".pos" suffix, `present_pos` with bare motor names.
+            goal_present_pos = {
+                key: (g_pos, present_pos[key.removesuffix(".pos")]) for key, g_pos in arm_goal_pos.items()
+            }
            arm_safe_goal_pos = ensure_safe_goal_position(goal_present_pos, self.config.max_relative_target)
            arm_goal_pos = arm_safe_goal_pos

@@ -77,11 +77,13 @@ class EEReferenceAndDelta(RobotActionProcessorStep):
    _command_when_disabled: np.ndarray | None = field(default=None, init=False, repr=False)

    def action(self, action: RobotAction) -> RobotAction:
-        observation = self.transition.get(TransitionKey.OBSERVATION).copy()
+        raw_observation = self.transition.get(TransitionKey.OBSERVATION)

-        if observation is None:
+        if raw_observation is None:
            raise ValueError("Joints observation is require for computing robot kinematics")

+        observation = raw_observation.copy()
+
        if self.use_ik_solution and "IK_solution" in self.transition.get(TransitionKey.COMPLEMENTARY_DATA):
            q_raw = self.transition.get(TransitionKey.COMPLEMENTARY_DATA)["IK_solution"]
        else:
@@ -311,10 +313,12 @@ class InverseKinematicsEEToJoints(RobotActionProcessorStep):
                "Missing required end-effector pose components: ee.x, ee.y, ee.z, ee.wx, ee.wy, ee.wz, ee.gripper_pos must all be present in action"
            )

-        observation = self.transition.get(TransitionKey.OBSERVATION).copy()
-        if observation is None:
+        raw_observation = self.transition.get(TransitionKey.OBSERVATION)
+        if raw_observation is None:
            raise ValueError("Joints observation is require for computing robot kinematics")

+        observation = raw_observation.copy()
+
        q_raw = np.array(
            [float(v) for k, v in observation.items() if isinstance(k, str) and k.endswith(".pos")],
            dtype=float,
@@ -391,13 +395,15 @@ class GripperVelocityToJoint(RobotActionProcessorStep):
    discrete_gripper: bool = False

    def action(self, action: RobotAction) -> RobotAction:
-        observation = self.transition.get(TransitionKey.OBSERVATION).copy()
+        raw_observation = self.transition.get(TransitionKey.OBSERVATION)

        gripper_vel = action.pop("ee.gripper_vel")

-        if observation is None:
+        if raw_observation is None:
            raise ValueError("Joints observation is require for computing robot kinematics")

+        observation = raw_observation.copy()
+
        q_raw = np.array(
            [float(v) for k, v in observation.items() if isinstance(k, str) and k.endswith(".pos")],
            dtype=float,
@@ -583,10 +589,12 @@ class InverseKinematicsRLStep(ProcessorStep):
                "Missing required end-effector pose components: ee.x, ee.y, ee.z, ee.wx, ee.wy, ee.wz, ee.gripper_pos must all be present in action"
            )

-        observation = new_transition.get(TransitionKey.OBSERVATION).copy()
-        if observation is None:
+        raw_observation = new_transition.get(TransitionKey.OBSERVATION)
+        if raw_observation is None:
            raise ValueError("Joints observation is require for computing robot kinematics")

+        observation = raw_observation.copy()
+
        q_raw = np.array(
            [float(v) for k, v in observation.items() if isinstance(k, str) and k.endswith(".pos")],
            dtype=float,
@@ -38,6 +38,11 @@ from .context import (
    RuntimeContext,
    build_rollout_context,
 )
+from .controller import (
+    LinkedEvent,
+    RolloutController,
+    RolloutEvent,
+)
 from .inference import (
    InferenceEngine,
    InferenceEngineConfig,
@@ -47,6 +52,11 @@ from .inference import (
    SyncInferenceEngine,
    create_inference_engine,
 )
+from .interactive import (
+    InteractiveCommand,
+    InteractiveSession,
+    parse_command,
+)
 from .strategies import (
    BaseStrategy,
    DAggerStrategy,
@@ -65,19 +75,24 @@ __all__ = [
    "DAggerStrategy",
    "DAggerStrategyConfig",
    "DatasetContext",
+    "EpisodicStrategy",
+    "EpisodicStrategyConfig",
    "HardwareContext",
    "HighlightStrategy",
    "HighlightStrategyConfig",
-    "EpisodicStrategy",
-    "EpisodicStrategyConfig",
    "InferenceEngine",
    "InferenceEngineConfig",
+    "InteractiveCommand",
+    "InteractiveSession",
+    "LinkedEvent",
    "PolicyContext",
    "ProcessorContext",
    "RTCInferenceConfig",
    "RTCInferenceEngine",
    "RolloutConfig",
    "RolloutContext",
+    "RolloutController",
+    "RolloutEvent",
    "RolloutStrategy",
    "RolloutStrategyConfig",
    "RuntimeContext",
@@ -88,4 +103,5 @@ __all__ = [
    "build_rollout_context",
    "create_inference_engine",
    "create_strategy",
+    "parse_command",
 ]
@@ -239,23 +239,18 @@ class RolloutConfig:
    # Runtime
    fps: float = 30.0
    duration: float = 0.0  # 0 = infinite (24/7 mode)
+    # Interactive session: control the rollout from stdin with chat-style
+    # commands (/start, /subtask <text>, /reset, /stop) while hardware and
+    # policy stay warm.  The robot does not move until /start is received,
+    # `/subtask` re-instructs the policy mid-run, and logs below ERROR are
+    # muted while the session runs so they don't interleave with the prompt.
+    # Supported with --strategy.type=base (no recording) and sentry
+    # (continuous recording; frames are labeled with the live task).
+    interactive: bool = False
    interpolation_multiplier: int = 1
-    # Safety net (opt-in): if any commanded joint's target differs from the robot's
-    # currently-measured position by more than this many units in a single control
-    # step, the action is treated as an unsafe jump — the robot is NOT commanded and
-    # the rollout stops. Units are the robot's raw `.pos` values (degrees for the
-    # SO-arms). Gripper joints are excluded (different unit/range). `None` disables it.
-    # Guards against chunk-splice / bad-chunk slams; a normal per-tick move at fps=30
-    # is only a few degrees, so a threshold like 20-30 catches slams without tripping
-    # on legitimate fast motion.
-    max_action_jump_deg: float | None = None
    device: str | None = None
    task: str = ""
    display_data: bool = False
-    # Also visualize model "extras" (e.g. a world model's imagined video) alongside observations.
-    # Off by default: requesting predictions forces per-chunk decoding on the control thread and only
-    # world-model policies produce anything. Implies display_data. Sync inference only.
-    display_extra_data: bool = False
    # Visualization backend used when display_data is True: "rerun" or "foxglove".
    display_mode: str = "rerun"
    # For "rerun": IP of a remote server to send to. For "foxglove": interface to bind the WebSocket
@@ -285,26 +280,6 @@ class RolloutConfig:

    def __post_init__(self):
        """Validate config invariants and load the policy config from ``--policy.path``."""
-        # --- Visualization validation ---
-        # Extra-data visualization piggybacks on the display_data path (backend init + telemetry
-        # logging are both gated on display_data), so enabling it implies display_data.
-        if self.display_extra_data and not self.display_data:
-            logger.info("display_extra_data=True implies display_data=True; enabling display_data")
-            self.display_data = True
-        # Only the sync engine surfaces intermediate predictions (RTC runs the policy in a background
-        # thread); warn and let it be ignored rather than fail.
-        if self.display_extra_data and not isinstance(self.inference, SyncInferenceConfig):
-            logger.warning(
-                "display_extra_data is only supported with sync inference (--inference.type=sync); "
-                "it will be ignored for inference type '%s'",
-                self.inference.type,
-            )
-
-        if self.max_action_jump_deg is not None and self.max_action_jump_deg <= 0:
-            raise ValueError(
-                f"max_action_jump_deg must be positive when set, got {self.max_action_jump_deg}"
-            )
-
        # --- Strategy-specific validation ---
        if isinstance(self.strategy, DAggerStrategyConfig) and self.teleop is None:
            raise ValueError("DAgger strategy requires --teleop.type to be set")
@@ -327,6 +302,17 @@ class RolloutConfig:
                "Base strategy does not record data. Use sentry, highlight, or dagger for recording."
            )

+        # Interactive mode drives strategy.run() in restartable segments and reads
+        # commands from stdin.  Base and sentry qualify: their run() loops keep no
+        # per-run terminal or dataset-finalization state.  The other recording
+        # strategies are excluded for now: they bind their own keyboard controls
+        # (which fight the command prompt for the terminal) and their run() loops
+        # finalize the dataset on exit, so they cannot be restarted.
+        if self.interactive and not isinstance(self.strategy, (BaseStrategyConfig, SentryStrategyConfig)):
+            raise ValueError(
+                f"--interactive=true supports --strategy.type=base or sentry (got '{self.strategy.type}')."
+            )
+
        # Sentry MUST use streaming encoding to avoid disk I/O blocking the control loop
        if (
            isinstance(self.strategy, SentryStrategyConfig)
@@ -45,6 +45,7 @@ from lerobot.processor import (
    make_default_processors,
    rename_stats,
 )
+from lerobot.processor.relative_action_processor import RelativeActionsProcessorStep
 from lerobot.robots import make_robot_from_config
 from lerobot.teleoperators import Teleoperator, make_teleoperator_from_config
 from lerobot.utils.feature_utils import combine_feature_dicts, hw_to_dataset_features
@@ -54,6 +55,7 @@ from .configs import BaseStrategyConfig, DAggerStrategyConfig, RolloutConfig
 from .inference import (
    InferenceEngine,
    RTCInferenceConfig,
+    SyncInferenceConfig,
    create_inference_engine,
 )
 from .inference.rtc import supports_rtc_inference
@@ -472,6 +474,15 @@ def build_rollout_context(
        },
    )

+    if isinstance(cfg.inference, SyncInferenceConfig) and any(
+        isinstance(step, RelativeActionsProcessorStep) and step.enabled
+        for step in getattr(preprocessor, "steps", ())
+    ):
+        raise NotImplementedError(
+            "SyncInferenceEngine does not support policies with relative actions for now."
+            "Use --inference.type=rtc or remove relative action processor steps from the policy pipeline."
+        )
+
    # --- 7. Inference strategy (needs policy + pre/post + hardware) --
    logger.info(
        "Creating inference engine (type=%s)...",
@@ -493,7 +504,6 @@ def build_rollout_context(
        use_torch_compile=torch_compile_active,
        compile_warmup_inferences=cfg.compile_warmup_inferences,
        shutdown_event=shutdown_event,
-        visualize_predictions=cfg.display_extra_data,
    )

    # --- 8. Assemble ---------------------------------------------------
@@ -0,0 +1,382 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Programmatic control of a rollout: start, pause, re-instruct, and stop a
+policy while hardware and policy stay connected and warm.
+
+:class:`RolloutController` is the embedding-friendly core of interactive
+rollouts.  It has no I/O of its own — no stdin, no printing, no log
+manipulation — so it can be driven from any application code: a CLI
+(:class:`lerobot.rollout.interactive.InteractiveSession` is exactly that), a
+network server, a voice front-end, or a notebook.
+
+Typical embedding::
+
+    from threading import Event, Thread
+    from lerobot.rollout import (
+        LinkedEvent,
+        RolloutController,
+        build_rollout_context,
+        create_strategy,
+    )
+
+    parent = Event()  # your application's shutdown signal
+    ctx = build_rollout_context(cfg, LinkedEvent(parent))
+    strategy = create_strategy(cfg.strategy)
+    strategy.setup(ctx)
+
+    controller = RolloutController(strategy, ctx)
+    serve_thread = Thread(target=controller.serve)
+    serve_thread.start()  # or call serve() on your main thread
+
+    controller.start()  # robot starts executing the policy
+    controller.set_task("grab the red cube")  # re-instruct mid-run
+    controller.reset()  # stop movement, return home, stay warm
+    controller.stop()  # end serve()
+
+    serve_thread.join()
+    strategy.teardown(ctx)  # teardown stays with the caller
+"""
+
+from __future__ import annotations
+
+import logging
+import time
+from collections.abc import Callable
+from enum import Enum
+from threading import Event, Lock
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from .context import RolloutContext
+    from .strategies import RolloutStrategy
+
+logger = logging.getLogger(__name__)
+
+
+class LinkedEvent(Event):
+    """A ``threading.Event`` whose ``is_set`` also reflects a parent event.
+
+    ``set``/``clear`` act only on the local flag, so a controller can raise
+    and clear its own segment-stop requests without masking (or accidentally
+    re-arming) the process-wide shutdown event carried by ``parent``.  Every
+    rollout strategy control loop polls ``ctx.runtime.shutdown_event.is_set()``,
+    so installing a ``LinkedEvent`` there makes the loops react both to
+    controller commands and to real shutdown signals.
+    """
+
+    _WAIT_SLICE_S = 0.05
+
+    def __init__(self, parent: Event) -> None:
+        super().__init__()
+        self.parent = parent
+
+    def is_set(self) -> bool:
+        return super().is_set() or self.parent.is_set()
+
+    def wait(self, timeout: float | None = None) -> bool:
+        """Wait for either the local or the parent flag.
+
+        The base ``Event.wait`` only watches the local flag, so poll in short
+        slices to also observe the parent.  Strategy loops only call
+        ``is_set()``; this coarse wait exists for API completeness.
+        """
+        deadline = None if timeout is None else time.perf_counter() + timeout
+        while not self.is_set():
+            remaining = None if deadline is None else deadline - time.perf_counter()
+            if remaining is not None and remaining <= 0:
+                return False
+            wait_slice = self._WAIT_SLICE_S if remaining is None else min(self._WAIT_SLICE_S, remaining)
+            super().wait(wait_slice)
+        return True
+
+
+class RolloutEvent(Enum):
+    """Lifecycle notifications emitted by :class:`RolloutController`.
+
+    All events are emitted on the thread running :meth:`RolloutController.serve`;
+    callbacks must be quick and must not call back into the controller's
+    blocking methods.
+    """
+
+    SEGMENT_STARTED = "segment_started"
+    """A control-loop segment is about to run (control state freshly reset)."""
+
+    SEGMENT_ENDED = "segment_ended"
+    """The segment returned on its own (e.g. ``--duration`` elapsed); the
+    controller is idle again and the robot is holding position."""
+
+    RESET_STARTED = "reset_started"
+    """A reset is being executed: inference paused, robot about to move home."""
+
+    RESET_DONE = "reset_done"
+    """The robot is back at its initial position, holding."""
+
+    RESET_SKIPPED = "reset_skipped"
+    """No initial position was captured; the robot holds its current pose."""
+
+    ENGINE_FAILED = "engine_failed"
+    """The inference engine hit an unrecoverable error; ``serve()`` is about
+    to return.  Read :attr:`RolloutController.failure_traceback` for details."""
+
+    STOPPED = "stopped"
+    """``serve()`` is returning (after :meth:`RolloutController.stop`, EOF of
+    the driving front-end, an engine failure, or a parent shutdown signal)."""
+
+
+class RolloutController:
+    """Drive a rollout strategy through thread-safe start/reset/stop/set_task calls.
+
+    The controller owns the outer lifecycle between ``strategy.setup(ctx)``
+    and ``strategy.teardown(ctx)`` (both stay with the caller): the robot is
+    idle until :meth:`start`, each run *segment* executes ``strategy.run(ctx)``
+    on the thread that called :meth:`serve` until interrupted or until the
+    strategy returns on its own (e.g. ``--duration`` elapsed).  :meth:`reset`
+    pauses the inference engine, returns the robot to its initial position,
+    and restores the launch task, while hardware and policy stay warm.
+    :meth:`stop` ends :meth:`serve` so the caller can run
+    ``strategy.teardown(ctx)``.
+
+    Requires ``ctx.runtime.shutdown_event`` to be a :class:`LinkedEvent`: the
+    controller sets the local flag to end a segment, and process shutdown
+    signals still propagate through the parent.  Build the context with
+    ``build_rollout_context(cfg, LinkedEvent(shutdown_event))``.
+
+    Thread safety: the control methods (:meth:`start`, :meth:`reset`,
+    :meth:`stop`, :meth:`set_task`) may be called from any thread and are
+    serialized by an internal lock, so calls issued in order from one thread
+    keep that order — e.g. a ``set_task`` right after a ``reset`` is not
+    clobbered by the reset's task restore.  Commands are last-write-wins:
+    ``reset`` and ``stop`` cancel a still-pending ``start`` so the robot
+    never starts moving after the caller's most recent command asked it not
+    to.  Events are emitted on the :meth:`serve` thread via ``on_event``.
+    """
+
+    _POLL_INTERVAL_S = 0.2
+
+    def __init__(
+        self,
+        strategy: RolloutStrategy,
+        ctx: RolloutContext,
+        on_event: Callable[[RolloutEvent], None] | None = None,
+    ) -> None:
+        stop_event = ctx.runtime.shutdown_event
+        if not isinstance(stop_event, LinkedEvent):
+            raise TypeError(
+                "RolloutController requires ctx.runtime.shutdown_event to be a LinkedEvent so "
+                "reset() can end a run segment without triggering process shutdown. Build the "
+                "rollout context with build_rollout_context(cfg, LinkedEvent(shutdown_event))."
+            )
+        self._strategy = strategy
+        self._ctx = ctx
+        self._segment_stop = stop_event
+        self._global_shutdown = stop_event.parent
+        self._on_event = on_event
+        # The instruction the rollout was launched with; reset() restores it.
+        self._initial_task = ctx.policy.inference.task
+
+        # Serializes the control methods so multi-writer task updates (e.g.
+        # reset()'s restore followed by a set_task()) keep their call order.
+        self._control_lock = Lock()
+
+        # Written by control methods (any thread), consumed by the serve loop.
+        self._start_requested = Event()
+        self._reset_requested = Event()
+        self._stop_requested = Event()
+        self._wake = Event()
+        self._running = Event()
+
+    # ------------------------------------------------------------------
+    # Introspection
+    # ------------------------------------------------------------------
+
+    @property
+    def task(self) -> str:
+        """The language instruction currently conditioning inference."""
+        return self._ctx.policy.inference.task
+
+    @property
+    def initial_task(self) -> str:
+        """The instruction the rollout was launched with (restored by :meth:`reset`)."""
+        return self._initial_task
+
+    @property
+    def running(self) -> bool:
+        """True while a control-loop segment is executing."""
+        return self._running.is_set()
+
+    @property
+    def failed(self) -> bool:
+        """True if the inference engine hit an unrecoverable error."""
+        return self._ctx.policy.inference.failed
+
+    @property
+    def failure_traceback(self) -> str | None:
+        """Formatted traceback of the engine failure, when :attr:`failed` is True."""
+        return self._ctx.policy.inference.failure_traceback
+
+    # ------------------------------------------------------------------
+    # Control methods (callable from any thread)
+    # ------------------------------------------------------------------
+
+    def start(self) -> bool:
+        """Request a control-loop segment.
+
+        Returns ``False`` when a segment is already running (the request is
+        ignored); ``True`` when the segment was scheduled.  The segment itself
+        executes on the :meth:`serve` thread.
+        """
+        with self._control_lock:
+            if self._running.is_set():
+                return False
+            self._start_requested.set()
+            self._wake.set()
+            return True
+
+    def reset(self) -> bool:
+        """Stop movement, return the robot to its initial position, restore the launch task.
+
+        Hardware and policy stay warm; call :meth:`start` to run again.
+        Returns ``True`` when the task was restored to the launch task (i.e.
+        it had been changed), ``False`` when it was already the launch task.
+        """
+        with self._control_lock:
+            # Last command wins: a start() still waiting to be serviced is
+            # cancelled so the robot never starts moving after the caller
+            # asked it not to.  Flag first, segment-stop second (see the
+            # ordering note in _run_segment).
+            self._start_requested.clear()
+            # Restore the task here, under the control lock, rather than in
+            # _reset_robot (which runs later, on the serve thread) so that a
+            # set_task() issued right after this reset() is not silently
+            # reverted by a deferred restore.
+            restored = self._ctx.policy.inference.set_task(self._initial_task)
+            self._reset_requested.set()
+            self._segment_stop.set()
+            self._wake.set()
+            return restored
+
+    def stop(self) -> None:
+        """End :meth:`serve`; the caller then runs ``strategy.teardown(ctx)``."""
+        with self._control_lock:
+            self._start_requested.clear()  # last command wins, see reset()
+            self._stop_requested.set()
+            self._segment_stop.set()
+            self._wake.set()
+
+    def set_task(self, task: str) -> bool:
+        """Change the instruction the policy follows, effective from the next inference.
+
+        Returns ``True`` when the value actually changed.  Safe to call while
+        a segment is running: the engine applies the switch on its own
+        inference thread (sync backends also drop actions precomputed under
+        the previous instruction).
+        """
+        with self._control_lock:
+            return self._ctx.policy.inference.set_task(task)
+
+    # ------------------------------------------------------------------
+    # Serve loop (blocks the calling thread)
+    # ------------------------------------------------------------------
+
+    def serve(self) -> None:
+        """Service control requests until :meth:`stop`, engine failure, or parent shutdown.
+
+        Blocks the calling thread; run segments execute here.  Emits
+        :class:`RolloutEvent` notifications through ``on_event``.
+        """
+        try:
+            while not self._global_shutdown.is_set():
+                if self._ctx.policy.inference.failed:
+                    self._emit(RolloutEvent.ENGINE_FAILED)
+                    break
+                if self._stop_requested.is_set():
+                    break
+                if self._reset_requested.is_set():
+                    self._reset_requested.clear()
+                    self._reset_robot()
+                    continue
+                if self._start_requested.is_set():
+                    # Consume the request and mark the segment running in one
+                    # atomic step: start() gates on _running, so a concurrent
+                    # start() is rejected for the entire startup sequence
+                    # (reset_control_state, SEGMENT_STARTED emission), not just
+                    # once strategy.run() begins — otherwise it could re-arm
+                    # _start_requested behind the running segment and the robot
+                    # would start again, uncommanded, when the segment ends.
+                    with self._control_lock:
+                        starting = self._start_requested.is_set()
+                        if starting:
+                            self._start_requested.clear()
+                            self._running.set()
+                    if starting:
+                        self._run_segment()
+                    continue
+                self._wake.wait(timeout=self._POLL_INTERVAL_S)
+                self._wake.clear()
+        finally:
+            self._emit(RolloutEvent.STOPPED)
+
+    def _run_segment(self) -> None:
+        """Execute one ``strategy.run`` segment until interrupted or finished.
+
+        The serve loop has already set ``_running`` (under the control lock),
+        so this method must clear it on every exit path.
+        """
+        engine = self._ctx.policy.inference
+        try:
+            # Clear the local flag *before* checking the request flags: control
+            # methods set their flag first and the segment-stop event second, so
+            # a reset() or stop() racing with this start() is either seen here or
+            # ends the freshly started loop on its first tick.
+            self._segment_stop.clear()
+            if (
+                self._stop_requested.is_set()
+                or self._reset_requested.is_set()
+                or self._global_shutdown.is_set()
+            ):
+                return
+            self._strategy.reset_control_state()
+            self._emit(RolloutEvent.SEGMENT_STARTED)
+            try:
+                self._strategy.run(self._ctx)
+            finally:
+                engine.pause()
+        finally:
+            self._running.clear()
+        if engine.failed:
+            return  # the serve loop emits ENGINE_FAILED and shuts down
+        if not (
+            self._stop_requested.is_set() or self._reset_requested.is_set() or self._global_shutdown.is_set()
+        ):
+            self._emit(RolloutEvent.SEGMENT_ENDED)
+
+    def _reset_robot(self) -> None:
+        """Pause inference and return the robot home (the task was restored by :meth:`reset`)."""
+        self._emit(RolloutEvent.RESET_STARTED)
+        self._ctx.policy.inference.pause()
+        if self._ctx.hardware.initial_position:
+            self._strategy.return_to_initial_position(self._ctx.hardware)
+            self._emit(RolloutEvent.RESET_DONE)
+        else:
+            logger.warning("No initial position captured — skipping the return move")
+            self._emit(RolloutEvent.RESET_SKIPPED)
+
+    def _emit(self, event: RolloutEvent) -> None:
+        if self._on_event is None:
+            return
+        try:
+            self._on_event(event)
+        except Exception:  # a broken observer must not kill the serve loop
+            logger.exception("Error in RolloutController event callback for %s", event)
@@ -22,9 +22,13 @@ or asynchronously in a background thread (RTC).
 from __future__ import annotations

 import abc
+import logging
+from threading import Lock

 import torch

+logger = logging.getLogger(__name__)
+

 class InferenceEngine(abc.ABC):
    """Abstract backend for producing actions during rollout.
@@ -47,12 +51,69 @@ class InferenceEngine(abc.ABC):
    backends always compute from ``obs_frame``; async backends ignore
    it (they receive observations via ``notify_observation``).

+    Task
+    ----
+    ``task`` / ``set_task`` hold the language instruction the policy is
+    conditioned on.  ``set_task`` is safe to call from any thread (the
+    interactive session's ``/subtask`` command calls it from its stdin
+    reader); subclasses pick the new value up on their own inference
+    thread via :meth:`_take_task`, so no policy state is ever mutated
+    across threads.
+
    Optional hooks
    --------------
    ``notify_observation`` / ``pause`` / ``resume`` have a no-op default
    so rollout strategies can invoke them unconditionally.
+
+    Subclasses must call ``super().__init__(task=...)``; the task holder
+    is set up there.
    """

+    def __init__(self, task: str = "") -> None:
+        self._task = task
+        self._task_changed = False
+        self._task_lock = Lock()
+
+    # ------------------------------------------------------------------
+    # Task (language instruction)
+    # ------------------------------------------------------------------
+
+    @property
+    def task(self) -> str:
+        """The language instruction currently conditioning inference."""
+        with self._task_lock:
+            return self._task
+
+    def set_task(self, task: str) -> bool:
+        """Set the instruction used from the next inference onwards.
+
+        Callable from any thread.  Returns ``True`` when the value
+        actually changed, so callers can report no-op switches.
+        """
+        with self._task_lock:
+            if task == self._task:
+                return False
+            previous, self._task = self._task, task
+            self._task_changed = True
+        logger.info("Task changed: '%s' -> '%s'", previous, task)
+        return True
+
+    def _take_task(self) -> tuple[str, bool]:
+        """Read the task and whether it changed since the last read.
+
+        Call from the thread that runs inference: the "changed" edge is
+        consumed here so the backend can drop actions precomputed under
+        the previous instruction before using the new one.
+        """
+        with self._task_lock:
+            changed, self._task_changed = self._task_changed, False
+            return self._task, changed
+
+    def _discard_task_change(self) -> None:
+        """Drop a pending task-change edge, e.g. from ``reset`` (state is already cleared)."""
+        with self._task_lock:
+            self._task_changed = False
+
    @abc.abstractmethod
    def start(self) -> None:
        """Initialise the backend."""
@@ -69,15 +130,6 @@ class InferenceEngine(abc.ABC):
    def get_action(self, obs_frame: dict | None) -> torch.Tensor | None:
        """Return the next action tensor, or ``None`` if unavailable."""

-    def get_intermediate_predictions(self) -> dict | None:
-        """Extra display-ready model outputs to visualize this tick, or ``None``.
-
-        Lets a backend surface a world model's intermediate predictions (e.g. imagined video
-        frames) into the rollout visualization path, keyed by ``"<datatype>.<name>"`` (mirroring
-        observation feature keys). Default: nothing extra.
-        """
-        return None
-
    def notify_observation(self, obs: dict) -> None:  # noqa: B027
        """Publish the latest processed observation.  Default: no-op."""

@@ -96,3 +148,8 @@ class InferenceEngine(abc.ABC):
    def failed(self) -> bool:
        """True if an unrecoverable error occurred in the backend."""
        return False
+
+    @property
+    def failure_traceback(self) -> str | None:
+        """Formatted traceback of the unrecoverable error, when ``failed`` is True."""
+        return None
@@ -95,7 +95,6 @@ def create_inference_engine(
    use_torch_compile: bool = False,
    compile_warmup_inferences: int = 2,
    shutdown_event: Event | None = None,
-    visualize_predictions: bool = False,
 ) -> InferenceEngine:
    """Instantiate the appropriate inference engine from a config object."""
    logger.info("Creating inference engine: %s", config.type)
@@ -109,7 +108,6 @@ def create_inference_engine(
            task=task,
            device=device,
            robot_type=robot_wrapper.robot_type,
-            visualize_predictions=visualize_predictions,
        )
    if isinstance(config, RTCInferenceConfig):
        return RTCInferenceEngine(
@@ -118,8 +116,7 @@ def create_inference_engine(
            postprocessor=postprocessor,
            robot_wrapper=robot_wrapper,
            rtc_config=config.rtc,
-            dataset_features=dataset_features,
-            ordered_action_keys=ordered_action_keys,
+            hw_features=hw_features,
            task=task,
            fps=fps,
            device=device,
@@ -35,13 +35,12 @@ import torch
 from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.policies.rtc import ActionQueue, LatencyTracker, reanchor_relative_rtc_prefix
 from lerobot.policies.rtc.configuration_rtc import RTCConfig
-from lerobot.policies.utils import make_robot_action, prepare_observation_for_inference
+from lerobot.policies.utils import prepare_observation_for_inference
 from lerobot.processor import (
    NormalizerProcessorStep,
    PolicyProcessorPipeline,
    RelativeActionsProcessorStep,
 )
-from lerobot.utils.constants import ACTION
 from lerobot.utils.feature_utils import build_dataset_frame

 from ..robot_wrapper import ThreadSafeRobot
@@ -82,28 +81,17 @@ def supports_rtc_inference(policy: PreTrainedPolicy) -> bool:


 def _normalize_prev_actions_length(prev_actions: torch.Tensor, target_steps: int) -> torch.Tensor:
-    """Pad or truncate RTC prefix actions to a fixed length for stable compiled inference.
-
-    Padding repeats the last real action ("hold") rather than filling with zeros. The RTC
-    guidance pulls the new chunk toward this prefix at the padded indices (they fall inside
-    the weighted region when the real leftover is shorter than ``target_steps``). A zero in
-    the model's normalized action space decodes to the dataset *mean* action — a nonzero
-    offset that yanks the spliced action toward a mean/neutral pose for one step, producing
-    an intermittent seam (e.g. 95 -> 103 -> 95). Holding the last real action keeps the
-    padded targets continuous with the prefix, so no fake target enters the guided region.
-    The fixed output length is preserved so ``torch.compile`` policies keep stable shapes.
-    """
+    """Pad or truncate RTC prefix actions to a fixed length for stable compiled inference."""
    if prev_actions.ndim != 2:
        raise ValueError(f"Expected 2D [T, A] tensor, got shape={tuple(prev_actions.shape)}")
-    steps, _ = prev_actions.shape
+    steps, action_dim = prev_actions.shape
    if steps == target_steps:
        return prev_actions
    if steps > target_steps:
        return prev_actions[:target_steps]
-    if steps == 0:
-        raise ValueError("Cannot pad an empty prefix: no last action to hold.")
-    hold = prev_actions[-1:].expand(target_steps - steps, -1)
-    return torch.cat([prev_actions, hold], dim=0)
+    padded = torch.zeros((target_steps, action_dim), dtype=prev_actions.dtype, device=prev_actions.device)
+    padded[:steps] = prev_actions
+    return padded


 # ---------------------------------------------------------------------------
@@ -127,8 +115,7 @@ class RTCInferenceEngine(InferenceEngine):
        postprocessor: PolicyProcessorPipeline,
        robot_wrapper: ThreadSafeRobot,
        rtc_config: RTCConfig,
-        dataset_features: dict,
-        ordered_action_keys: list[str],
+        hw_features: dict,
        task: str,
        fps: float,
        device: str | None,
@@ -137,37 +124,13 @@ class RTCInferenceEngine(InferenceEngine):
        rtc_queue_threshold: int = 30,
        shutdown_event: Event | None = None,
    ) -> None:
+        super().__init__(task=task)
        self._policy = policy
        self._preprocessor = preprocessor
        self._postprocessor = postprocessor
        self._robot = robot_wrapper
        self._rtc_config = rtc_config
-        # Build observations with the SAME feature spec sync uses (post
-        # `robot_observation_processor`), not the raw-hardware spec. `build_dataset_frame`
-        # orders `observation.state` by this spec's `names`; using the raw-hardware order
-        # here (as before) desynced the state vector from sync whenever the observation
-        # processor reorders/renames state keys, corrupting both normalization and the
-        # relative-action anchor. The `prefix="observation"` filter ignores the action
-        # entries in the combined dict.
-        self._obs_features = dataset_features
-        # The model emits actions in `dataset_features[ACTION]` order (the order it was
-        # trained on); the robot expects them in `ordered_action_keys` order. Sync remaps
-        # by NAME via `make_robot_action` + reindex (sync.py) before returning; RTC must do
-        # the SAME, otherwise the engine-agnostic strategy (`send_next_action`) maps the raw
-        # model-order tensor onto `ordered_action_keys` positionally and mis-assigns joints
-        # whenever the two orders differ — a per-joint permutation that drives the arm wrong.
-        self._ordered_action_keys = ordered_action_keys
-        state_ft = dataset_features.get("observation.state")
-        if state_ft is not None:
-            logger.info("RTC observation.state layout: %s", state_ft.get("names"))
-        action_ft = dataset_features.get(ACTION)
-        if action_ft is not None:
-            logger.info(
-                "RTC action layout: model/dataset=%s -> robot=%s",
-                action_ft.get("names"),
-                self._ordered_action_keys,
-            )
-        self._task = task
+        self._hw_features = hw_features
        self._fps = fps
        self._device = device or "cpu"
        self._use_torch_compile = use_torch_compile
@@ -177,10 +140,14 @@ class RTCInferenceEngine(InferenceEngine):
        self._action_queue: ActionQueue | None = None
        self._obs_holder: dict[str, Any] = {}
        self._obs_lock = Lock()
+        # Bumped by reset() (under _obs_lock) so chunks whose inference started
+        # before a reset are discarded instead of merged into the fresh queue.
+        self._reset_epoch = 0
        self._policy_active = Event()
        self._compile_warmup_done = Event()
        self._shutdown_event = Event()
        self._rtc_error = Event()
+        self._failure_traceback: str | None = None
        self._global_shutdown_event = shutdown_event
        self._rtc_thread: Thread | None = None

@@ -227,6 +194,15 @@ class RTCInferenceEngine(InferenceEngine):
        """True if the RTC background thread exited due to an unrecoverable error."""
        return self._rtc_error.is_set()

+    @property
+    def failure_traceback(self) -> str | None:
+        """Traceback captured when the RTC thread died (see ``failed``).
+
+        Kept on the engine so consumers that mute console logging (the
+        interactive session) can still surface the fatal error.
+        """
+        return self._failure_traceback
+
    @property
    def action_queue(self) -> ActionQueue | None:
        """The shared action queue between the RTC thread and the main loop."""
@@ -272,32 +248,39 @@ class RTCInferenceEngine(InferenceEngine):
        self._policy_active.set()

    def reset(self) -> None:
-        """Reset the policy, processors, and action queue."""
+        """Reset the policy, processors, and action queue.
+
+        Call while the engine is paused (both DAgger transitions and the
+        interactive session do): the RTC thread may still be finishing an
+        inference started before the pause, so ``reset`` also drops the last
+        published observation — it can be arbitrarily stale by the time the
+        engine resumes (e.g. the robot was returned to its initial position
+        in the meantime), and a chunk computed from it would jerk the robot
+        toward the old pose — and bumps the reset epoch so any in-flight
+        chunk is discarded instead of merged into the cleared queue.
+        """
        logger.info("Resetting RTC inference state (policy + processors + queue)")
        self._policy.reset()
        self._preprocessor.reset()
        self._postprocessor.reset()
        if self._action_queue is not None:
            self._action_queue.clear()
+        with self._obs_lock:
+            self._obs_holder["obs"] = None
+            self._reset_epoch += 1
+        # The queue was just cleared, so a pending task change has nothing
+        # stale left to blend against.
+        self._discard_task_change()

    # ------------------------------------------------------------------
    # Action production (called from main thread)
    # ------------------------------------------------------------------

    def get_action(self, obs_frame: dict | None) -> torch.Tensor | None:
-        """Pop the next action from the RTC queue (ignores ``obs_frame``).
-
-        The queued action is in the model's ``dataset_features[ACTION]`` order; remap it
-        by NAME into ``ordered_action_keys`` order before returning, so the engine-agnostic
-        strategy maps values onto the correct joints. Mirrors ``SyncInferenceEngine.get_action``.
-        """
+        """Pop the next action from the RTC queue (ignores ``obs_frame``)."""
        if self._action_queue is None:
            return None
-        action = self._action_queue.get()
-        if action is None:
-            return None
-        action_dict = make_robot_action(action, self._obs_features)
-        return torch.tensor([action_dict[k] for k in self._ordered_action_keys])
+        return self._action_queue.get()

    def notify_observation(self, obs: dict) -> None:
        """Publish the latest observation for the RTC thread to consume."""
@@ -316,8 +299,6 @@ class RTCInferenceEngine(InferenceEngine):
            policy_device = torch.device(self._device)

            warmup_required = max(1, self._compile_warmup_inferences) if self._use_torch_compile else 0
-            # exclude the first N inferences from the latency tracker to avoid cold-start spikes
-            latency_warmup_required = max(1, warmup_required)
            inference_count = 0
            consecutive_errors = 0

@@ -329,6 +310,7 @@ class RTCInferenceEngine(InferenceEngine):
                queue = self._action_queue
                with self._obs_lock:
                    obs = self._obs_holder.get("obs")
+                    epoch_before = self._reset_epoch
                if queue is None or obs is None:
                    time.sleep(_RTC_IDLE_SLEEP_S)
                    continue
@@ -339,14 +321,27 @@ class RTCInferenceEngine(InferenceEngine):
                        idx_before = queue.get_action_index()
                        prev_actions = queue.get_left_over()

-                        latency = latency_tracker.p95()
+                        latency = latency_tracker.max()
                        delay = math.ceil(latency / time_per_chunk) if latency else 0

-                        obs_batch = build_dataset_frame(self._obs_features, obs, prefix="observation")
+                        task, task_changed = self._take_task()
+                        if task_changed:
+                            # No queue flush on purpose: dropping the queued
+                            # actions would leave the robot without commands for
+                            # a full inference latency.  With RTC blending on
+                            # (the default) this chunk — already conditioned on
+                            # the new instruction — is merged over the previous
+                            # chunk's leftover prefix, so the switch lands within
+                            # one inference and the transition stays continuous.
+                            # With blending disabled the queue drains first, so
+                            # it lands up to one chunk later.
+                            logger.info("Task changed to '%s' — applied from this chunk on", task)
+
+                        obs_batch = build_dataset_frame(self._hw_features, obs, prefix="observation")
                        obs_batch = prepare_observation_for_inference(
-                            obs_batch, policy_device, self._task, self._robot.robot_type
+                            obs_batch, policy_device, task, self._robot.robot_type
                        )
-                        obs_batch["task"] = [self._task]
+                        obs_batch["task"] = [task]

                        preprocessed = self._preprocessor(obs_batch)

@@ -382,13 +377,17 @@ class RTCInferenceEngine(InferenceEngine):
                        inference_count += 1
                        consecutive_errors = 0
                        is_warmup = self._use_torch_compile and inference_count <= warmup_required
-                        # Ignore the first N inferences for latency tracking to avoid cold-start spikes
-                        if inference_count <= latency_warmup_required:
+                        if is_warmup:
                            latency_tracker.reset()
                        else:
                            latency_tracker.add(new_latency)

-                        queue.merge(original, processed, new_delay, idx_before)
+                        with self._obs_lock:
+                            epoch_unchanged = epoch_before == self._reset_epoch
+                        if epoch_unchanged:
+                            queue.merge(original, processed, new_delay, idx_before)
+                        else:
+                            logger.info("Discarding action chunk computed before an engine reset")

                        if (
                            is_warmup
@@ -417,8 +416,9 @@ class RTCInferenceEngine(InferenceEngine):
                    time.sleep(_RTC_IDLE_SLEEP_S)

        except Exception as e:
+            self._failure_traceback = traceback.format_exc()
            logger.error("Fatal error in RTC thread: %s", e)
-            logger.error(traceback.format_exc())
+            logger.error(self._failure_traceback)
            self._rtc_error.set()
            # Unblock any warmup waiters so the main loop doesn't spin forever
            self._compile_warmup_done.set()
@@ -22,23 +22,28 @@ from copy import copy

 import torch

-from lerobot.policies.pretrained import PreTrainedPolicy, unpack_action_output
+from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.policies.utils import make_robot_action, prepare_observation_for_inference
-from lerobot.processor import PolicyProcessorPipeline, RelativeActionsProcessorStep
+from lerobot.processor import PolicyProcessorPipeline

 from .base import InferenceEngine

 logger = logging.getLogger(__name__)


-# Relative-action support: a predicted chunk of offsets is anchored to the robot
-# state at prediction time, but the sync engine reruns the pre/post pipeline every
-# tick, so ``RelativeActionsProcessorStep`` would re-anchor cached actions to the
-# current (moved) state and drift through the chunk. We pin the anchor per chunk:
-# a probe on the policy's public ``predict_action_chunk`` flags the ticks that
-# predict a fresh chunk; on the others the engine restores the anchor the relative
-# step overwrote. ``select_action`` stays on the hot path, so per-tick side effects
-# (e.g. LingBot-VA keyframe feedback) are preserved.
+# TODO(Steven): support relative-action policies.  The per-tick flow refreshes
+# ``RelativeActionsProcessorStep._last_state`` every call, so cached chunk
+# actions popped on later ticks get reanchored to the *current* robot state and
+# absolute targets drift through the chunk.  Relative-action policies are
+# rejected at context-build time today; RTC postprocesses the whole chunk and
+# is unaffected.
+#
+# Candidate fix: drive the policy via ``predict_action_chunk`` and serve a
+# local FIFO of postprocessed actions.  Eliminates drift by construction and
+# saves per-tick pre/post work, but bypasses ``select_action`` — needs
+# fallbacks for SAC (raises), ACT temporal ensembling (ensembler lives in
+# ``select_action``), and Diffusion-family (obs-history queues populated as a
+# side effect of ``select_action``).


 class SyncInferenceEngine(InferenceEngine):
@@ -59,55 +64,19 @@ class SyncInferenceEngine(InferenceEngine):
        task: str,
        device: str | None,
        robot_type: str,
-        visualize_predictions: bool = False,
    ) -> None:
+        super().__init__(task=task)
        self._policy = policy
        self._preprocessor = preprocessor
        self._postprocessor = postprocessor
        self._dataset_features = dataset_features
        self._ordered_action_keys = ordered_action_keys
-        self._task = task
        self._device = torch.device(device or "cpu")
        self._robot_type = robot_type
-
-        # Find an enabled RelativeActionsProcessorStep to pin its anchor per chunk
-        # (see module comment), mirroring the RTC engine.
-        self._relative_step = next(
-            (
-                s
-                for s in getattr(preprocessor, "steps", ())
-                if isinstance(s, RelativeActionsProcessorStep) and s.enabled
-            ),
-            None,
-        )
-        # Set by the probe for the current tick / ever, respectively.
-        self._chunk_predicted = False
-        self._ever_predicted_chunk = False
-        self._original_predict_action_chunk = None  # set while the probe is installed
-        if self._relative_step is not None:
-            # ``action_names`` is optional on the step; fill it lazily from the
-            # policy/dataset so the relative<->absolute mask is built correctly. This is
-            # a deliberate engine->step side effect (the step is configured by its consumer).
-            if self._relative_step.action_names is None:
-                cfg_names = getattr(policy.config, "action_feature_names", None)
-                self._relative_step.action_names = list(cfg_names) if cfg_names else list(ordered_action_keys)
-            self._install_chunk_probe()
-            logger.info("Relative actions enabled: chunk anchor pinned per predicted chunk")
-
-        # Intermediate-prediction visualization (e.g. a world model's imagined video). When on,
-        # ``get_action`` requests predictions and keeps the current chunk's frame stacks; a playhead
-        # (``get_intermediate_predictions``) advances one step per tick, paced across the chunk's tick
-        # span so the imagined clip stays wall-clock aligned with execution.
-        self._visualize_predictions = visualize_predictions
-        self._pred_stacks: dict = {}  # key -> [T, H, W, 3] frame stack for the current chunk
-        self._pred_cursor = 0  # ticks elapsed since the current chunk's frames arrived
-        self._ticks_per_chunk = getattr(getattr(policy, "config", None), "chunk_size", None)
-
        logger.info(
-            "SyncInferenceEngine initialized (device=%s, action_keys=%d, visualize_predictions=%s)",
+            "SyncInferenceEngine initialized (device=%s, action_keys=%d)",
            self._device,
            len(ordered_action_keys),
-            self._visualize_predictions,
        )

    def start(self) -> None:
@@ -116,11 +85,6 @@ class SyncInferenceEngine(InferenceEngine):

    def stop(self) -> None:
        """No background resources to stop."""
-        # Undo the probe so the policy object isn't left permanently patched
-        # (it may outlive this engine or be reused by another).
-        if self._original_predict_action_chunk is not None:
-            self._policy.predict_action_chunk = self._original_predict_action_chunk
-            self._original_predict_action_chunk = None
        logger.info("SyncInferenceEngine stopped")

    def reset(self) -> None:
@@ -129,54 +93,9 @@ class SyncInferenceEngine(InferenceEngine):
        self._policy.reset()
        self._preprocessor.reset()
        self._postprocessor.reset()
-        # New episode: the next tick predicts a fresh chunk and re-anchors.
-        self._chunk_predicted = False
-        self._ever_predicted_chunk = False
-        self._pred_stacks = {}
-        self._pred_cursor = 0
-
-    def _install_chunk_probe(self) -> None:
-        """Wrap the policy's public ``predict_action_chunk`` so we learn which ticks
-        predict a fresh chunk (when the anchor must advance) without introspecting any
-        private action queue. Chunking policies call it from ``select_action``.
-
-        Wraps whatever callable is currently bound (e.g. an already-``torch.compile``d
-        one, since ``build_rollout_context`` compiles before building the engine); undone
-        in ``stop()``."""
-        self._original_predict_action_chunk = self._policy.predict_action_chunk
-        inner = self._original_predict_action_chunk
-
-        def probe(*args, **kwargs):
-            self._chunk_predicted = True
-            self._ever_predicted_chunk = True
-            return inner(*args, **kwargs)
-
-        self._policy.predict_action_chunk = probe
-
-    def get_intermediate_predictions(self) -> dict | None:
-        """Serve one imagined frame per key for this tick, advancing the playhead.
-
-        Maps the current chunk's ``T`` decoded frames onto its ``ticks_per_chunk`` control ticks so
-        the imagined video plays back in step with execution (falls back to one frame/tick, clamped,
-        when the chunk's tick span is unknown). Returns ``None`` until a chunk with frames arrives.
-        """
-        if not self._pred_stacks:
-            return None
-        tick = self._pred_cursor
-        span = self._ticks_per_chunk
-        out: dict = {}
-        for key, stack in self._pred_stacks.items():
-            n = len(stack)
-            if n == 0:
-                continue
-            idx = round(tick / (span - 1) * (n - 1)) if span and span > 1 else tick
-            idx = min(max(idx, 0), n - 1)
-            frame = stack[idx]
-            if hasattr(frame, "detach"):
-                frame = frame.detach().cpu().numpy()
-            out[key] = frame
-        self._pred_cursor += 1
-        return out or None
+        # The policy was just reset, so a pending task change has nothing
+        # stale left to flush.
+        self._discard_task_change()

    def get_action(self, obs_frame: dict | None) -> torch.Tensor | None:
        """Run the full inference pipeline on ``obs_frame`` and return an action tensor."""
@@ -191,34 +110,22 @@ class SyncInferenceEngine(InferenceEngine):
            if self._device.type == "cuda" and self._policy.config.use_amp
            else nullcontext()
        )
-        # Snapshot the chunk anchor before the preprocessor overwrites it with this
-        # tick's state; restore it below if this tick only served a cached action.
-        # ``clone`` so the snapshot survives even if the cached tensor is ever mutated
-        # in place (today it is only rebound, but the copy is cheap for a state vector).
-        anchor_before = None
-        if self._relative_step is not None:
-            cached = self._relative_step.get_cached_state()
-            anchor_before = cached.clone() if cached is not None else None
-        self._chunk_predicted = False
+        task, task_changed = self._take_task()
        with torch.inference_mode(), autocast_ctx:
-            observation = prepare_observation_for_inference(
-                observation, self._device, self._task, self._robot_type
-            )
+            if task_changed:
+                # Chunking policies serve actions from an internal queue filled
+                # under the previous instruction (up to chunk_size ticks of stale
+                # behavior), so drop them and let the new instruction take effect
+                # on this very tick.  Deliberately narrower than ``policy.reset``:
+                # observation history and other episode state are kept, so a
+                # policy that conditions on them (and one that ignores the task
+                # entirely) sees no discontinuity.  Safe to mutate here — this is
+                # the thread that calls ``select_action``.
+                logger.info("Task changed to '%s' — dropping precomputed actions", task)
+                self._policy.drop_queued_actions()
+            observation = prepare_observation_for_inference(observation, self._device, task, self._robot_type)
            observation = self._preprocessor(observation)
-            if self._visualize_predictions:
-                action, predictions = unpack_action_output(
-                    self._policy.select_action(observation, return_intermediate_predictions=True)
-                )
-                if predictions:
-                    # A fresh chunk was predicted this tick — store its frame stacks and restart the playhead.
-                    self._pred_stacks = predictions
-                    self._pred_cursor = 0
-            else:
-                action = self._policy.select_action(observation)
-            # Hold the anchor only for a chunking policy serving a cached action this
-            # tick; policies that never chunk or that recomputed keep refreshing.
-            if self._relative_step is not None and self._ever_predicted_chunk and not self._chunk_predicted:
-                self._relative_step.set_cached_state(anchor_before)
+            action = self._policy.select_action(observation)
            action = self._postprocessor(action)
        action_tensor = action.squeeze(0).cpu()

--- a/Show More
+++ b/Show More