chore(dataset): add check dataset shape

Declaring a numeric feature with `shape=(0,)` crashed `save_episode()` in two
distinct places, leaving `dtype: "string"` as the only (type-lossy) workaround:

  - `compute_episode_stats` -> `RunningQuantileStats.update` reshaped a size-0
    array, raising "ValueError: cannot reshape array of size 0 into shape (0)".
  - `get_hf_features_from_features` mapped it to a fixed-size Arrow list of
    length 0, which pyarrow rejects ("list_size needs to be a strict positive
    integer").

The issue only reported the first error; the second surfaces once the first is
fixed. This change handles both:

  - Skip zero-width features during episode stats, exactly as string/language
    features are already skipped.
  - Store 1-D zero-width features as a variable-length sequence (length=-1) so
    each per-frame value is simply an empty list.

Adds a unit test (stats layer) and an integration test that records, saves, and
reads back a zero-width feature, asserting it round-trips as an empty vector and
is excluded from stats.

Fixes #3654

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

DOCS_AUDIT.md

@@ -1,113 +0,0 @@
-# LeRobot Docs Audit
-
-**Status:** Fresh baseline for a new docs redesign effort · **Date:** 2026-07-27
-**Supersedes:** the June 2026 `DOCS_REDESIGN.md` proposal (lived on the unmerged, now-stale `docs/complete-docs-redesign` branch — abandoned; that branch is ~67k lines behind `main` and should not be resurrected). This document re-audits `docs/source/` from scratch against current `main`, since the docs tree grew substantially in the ~6 weeks between the two audits (many new policies, robots, and benchmarks landed).
-
----
-
-## Snapshot (current state, in numbers)
-
-- **IA**: 15 top-level `_toctree.yml` sections, 81 `local:` entries, 0 broken toctree links.
-- **Files**: 97 total files in `docs/source/` = 80 real `.mdx` pages + 16 orphaned `policy_*_README.md` stubs (up from 14 in June) + 1 `contributing.md` (a genuine symlink to root `CONTRIBUTING.md`).
-- **Growth**: `.mdx` pages grew 49 → 80 in 6 months (+63%, ~5.2 pages/month), arriving in **bursts against a shared, hand-edited `_toctree.yml`** — e.g. 6 benchmark pages landed via 6 separate PRs within 48 hours (Apr 2026); 3 policy pages within 3 days (Jul 2026).
-- **CLI docs coverage improved sharply since June**: 17 of 18 `lerobot-*` commands now documented (only `lerobot-info`, a diagnostics helper, is not) — the June baseline was 6/17 undocumented.
-- **Decision content**: zero comparison/decision pages exist in `docs/source/` for Policies (15 options), Robots (11), Benchmarks (10), or Teleoperators — this equivalent content already exists, fully written, in root `AGENT_GUIDE.md` §6, but was never ported or linked in.
-- **Teleoperators**: dedicated toctree section has only 2 pages vs. 16-17 registered teleoperator types in `src/`, but all types do get at least one doc mention somewhere (usually buried inside a robot's hardware page) — a discoverability gap, not an absence of content.
-- **Orphan policy READMEs**: of the 16, 3 (Diffusion, TD-MPC, VQ-BeT) have **zero real published documentation anywhere** — the orphan stub is their only artifact. 13 duplicate an actively-maintained `.mdx` page under a different name. One (MolmoAct2) is proven to have silently rotted: a 39-line pre-refactor stub sits next to a 495-line maintained page that diverged 481 lines ago.
-- **Root `README.md` itself has drifted**: its "Supported Hardware" line omits SO-101 (the flagship robot) entirely; its "SoTA Models" table links several policies to thin orphan stubs instead of the richer tutorial pages that exist for them.
-- **4 concrete broken/stale copy-paste commands** verified in published pages (`act.mdx`, `il_robots.mdx` ×2 spots, `lerobot-dataset-v3.mdx`, `openarm.mdx` — the last added as recently as Jan 2026, so a fresh miss, not just aging drift).
-- **Process/CI**: no CODEOWNERS, no `_redirects.yml`, no doc-consistency or toctree-completeness check anywhere; the docs CI workflow triggers only on `paths: docs/**`, so code-only PRs get zero automated docs signal (empirically responsible for the one real lag found — OMX's docs page, 41 days after its code). Despite this, 18 of 19 spot-checked new integrations (EVO1, FastWAM, LingBot-VA, MolmoAct2, RTC, VLA-JEPA, X-VLA, WallOSS, Damiao, Hope Jr, OpenArm, reBot B601, Reachy 2, Unitree G1, RoboCasa365, VLABench, IsaacLab Arena, etc.) shipped docs in the *same commit* as the feature — organic discipline is currently strong.
-- **No content template** exists for policy or robot doc pages: heading sets/length vary 80-528 lines with almost no shared vocabulary; by contrast, benchmarks *do* have a written template (`adding_benchmarks.mdx`) and the resulting 8 pages are visibly consistent — direct proof a template is what produces consistency here.
-
----
-
-## Strengths
-
-1. **Organic contributor discipline is currently much better than the June audit's numbers implied.** 18/19 spot-checked new policies/robots/benchmarks shipped docs in the same PR as the code; 17/18 CLI commands are documented; all teleoperator types get at least one mention. The underlying practice is healthier than the structural numbers (orphans, thin nav sections) suggest — the real problem is increasingly discoverability and consistency, not absence of effort.
-2. **The doc-builder CI pipeline (PR previews, main build, versioned release builds) is solid, standard infrastructure already in place** — nothing bespoke to build or maintain, and it's the same tooling Transformers uses.
-3. **A working, provably-effective template pattern already exists for one catalog (benchmarks)** — `adding_benchmarks.mdx`'s checklist + "at a glance" table produces 8 structurally consistent pages. This is the single best evidence in the repo for what fixes the policy/robot template gap, and it needs no new invention, just extension.
-4. **Genuine single-source-of-truth patterns already exist and work**: `docs/source/contributing.md` is a real symlink to root `CONTRIBUTING.md` (verified via `ls -la`/`find -type l`). Most policy `src/README.md` files are also symlinked into `docs/source/policy_*_README.md`, eliminating drift at that layer (though the docs-side stub itself remains an unlinked orphan — see weaknesses).
-5. **Zero broken toctree links**, and specific pages are genuinely high quality: `cheat-sheet.mdx` (spot-checked line-by-line against source, fully accurate), `bring_your_own_policies.mdx` (clear scope, working PR checklist), `installation.mdx`'s tables/OS tabs, and the SO-100/SO-101 hardware pages.
-6. **Root `README.md` already contains a usable taxonomy for policies** (Imitation Learning / RL / VLAs / World Models / Reward Models) that the flat 15-item Policies sidebar never adopted — the raw material for a better IA already exists in-repo.
-
----
-
-## Weaknesses
-
-### High
-- **No comparison/decision-guidance page for any multi-option category** (Policies=15, Robots=11, Benchmarks=10, Teleoperators) — the exact content (decision rules, profiling snapshot) already exists in `AGENT_GUIDE.md` §6 but was never ported into `docs/source`.
-- **The landing page (`index.mdx`) provides zero navigation** — 23 lines of marketing copy and a Discord link, no path to installation, cheat-sheet, or audience-specific guidance.
-- **Teleoperators' dedicated nav section (2 pages) badly undersells real coverage** (16-17 implementations, all mentioned somewhere) — a new user browsing that sidebar section would wrongly conclude only phone/Isaac teleop exist. Content exists; it's a pure discoverability failure.
-- **Multiple concrete broken/stale CLI examples in published, frequently copy-pasted pages**: `act.mdx` tells readers to eval with `lerobot-record` while the shown code uses `lerobot-rollout`; an identical shell-breaking snippet (`\`-terminated comment swallowing a continuation) is duplicated in `il_robots.mdx` and `lerobot-dataset-v3.mdx`; `il_robots.mdx:421` uses the dead `--control.push_to_hub` flag namespace; `openarm.mdx:208-210` (added Jan 2026) gives a `lerobot-record` example with flat pre-draccus flags that don't exist on the current config classes.
-- **3 shipping policies (Diffusion, TD-MPC, VQ-BeT) have zero real published documentation** — only a thin, toctree-unreferenced stub exists for each.
-- **Root `README.md` has drifted as the project scaled**: its Supported Hardware line omits SO-101 entirely; its SoTA Models table links ACT/Diffusion/VQ-BeT/Multitask-DiT/TDMPC/GR00T/SmolVLA to thin orphan stubs instead of the richer tutorial pages that exist for several of them.
-- **No content template for policy/robot doc pages, unlike benchmarks** — heading sets and length vary 80-528 lines with almost no shared vocabulary (`molmoact2.mdx` has no Overview section at all; `smolvla.mdx` has no Overview/Architecture/Citation/License and reads as pure tutorial).
-- **Docs CI triggers only on `paths: docs/**`** — code-only PRs get zero automated docs signal; this is the mechanism directly responsible for the one measured lag (OMX's docs page shipped 41 days after its code, in a separate PR).
-- **Proof that orphan-stub drift isn't hypothetical**: `policy_molmoact2_README.md` is a stale 39-line pre-refactor snapshot sitting 481 lines behind the maintained 495-line page it once mirrored — nothing caught this silently rotting.
-
-### Medium
-- 16 orphaned `policy_*_README.md` files unreferenced by the toctree; 13 are pure duplicate cruft shadowing a maintained same-topic `.mdx` page, creating a false "two files to keep in sync" impression for contributors.
-- No `_redirects.yml` anywhere — future cleanup of the orphans/renames has no 404 safety net.
-- `installation.mdx` ends on a dangling forward-reference ("follow the link below to use LeRobot with your robot") — no link follows.
-- `cheat-sheet.mdx`'s "Policy Types" line is stale (`act, diffusion, smolvla, pi05`) against the current 15-entry Policies section, and lists `diffusion`, which has no working toctree page at all.
-- "Tutorials" (10 pages) mixes beginner, contributor, and RL-researcher content in one flat, unlabeled list.
-- The EnvHub feature family is split inconsistently across two unrelated sections (`envhub.mdx`/`envhub_leisaac.mdx` under Simulation, `envhub_isaaclab_arena.mdx` under Benchmarks) despite identical framing/opening text.
-- Admonition syntax is split GFM `[!NOTE]` vs. doc-builder `<Tip>` across sampled pages; since the site actually builds with `hf-doc-builder`, the majority pattern may not render as a styled callout at all (plausible, not independently render-verified).
-- No CODEOWNERS file anywhere — zero designated reviewer routing for `docs/source`.
-- `CONTRIBUTING.md` never links the good in-repo extension guides (`adding_benchmarks.mdx`, `bring_your_own_policies.mdx`, `integrate_hardware.mdx`) and states no docs-required policy.
-- The "docs required" PR checklist convention is applied inconsistently: Policies and Benchmarks have an explicit required-docs checklist; `integrate_hardware.mdx` (Robots/Teleoperators) has none — a plausible root cause of the Teleoperators gap above.
-- Growth repeatedly stacks simultaneous PRs against one shared, hand-edited `_toctree.yml` (6 benchmark PRs in 48h; 3 policy PRs in 3 days) — a merge/oversight risk even though no damage from it was found yet.
-- `policy_sarm_README.md` has no source of truth left to sync against at all (SARM moved `policies/`→`rewards/`, no README carried over) — pure abandoned cruft next to the real 593-line `sarm.mdx`.
-- No automated check anywhere for toctree completeness or README/mdx sync — the only CI gate is whether the doc-builder build succeeds, which doesn't catch missing coverage.
-
-### Low
-- No dedicated "Motors" section despite `motors/` being named as a distinct hardware layer in `CLAUDE.md`; `feetech.mdx`/`damiao.mdx` sit in a catch-all "Resources" section instead.
-- "Sensors" is a single-page top-level nav section (`cameras.mdx` only).
-- Policies (15), Benchmarks (10), and Robots (11) are flat, ungrouped lists with no internal sub-headings, despite `README.md` already having a usable taxonomy that could be reused.
-- "SO-101" vs. "SO101" naming is inconsistent across ~15 files, and even within a single file (`so100.mdx` uses both).
-- `hilserl.mdx` (950 lines) and `il_robots.mdx` (638 lines) each mix tutorial, reference, and troubleshooting content in one long page.
-- `reachy2_camera` has zero doc mentions anywhere; the `zmq` camera backend is documented only incidentally inside `unitree_g1.mdx` rather than centrally in `cameras.mdx`.
-- The PR template's "Documentation updated" line is a self-reported, unenforced checkbox.
-
----
-
-## Recommendations
-
-### 1. Quick wins (small effort, ship this week, no maintainer proposal needed)
-- Fix the 4 verified broken commands: `act.mdx` (`lerobot-record`→`lerobot-rollout`), the duplicated shell-breaking snippet in `il_robots.mdx` + `lerobot-dataset-v3.mdx`, `il_robots.mdx`'s `--control.push_to_hub`→`--dataset.push_to_hub`, and `openarm.mdx`'s invalid flat-flag record example.
-- Delete the 3 fully-orphaned policy stubs (Diffusion, TD-MPC, VQ-BeT) and give each a real toctree page by promoting the existing `src/` README content.
-- Delete the remaining 13 duplicate orphan stubs and the dead `policy_sarm_README.md`.
-- Fix `installation.mdx`'s dangling ending and `cheat-sheet.mdx`'s stale policy-type list.
-- Fix root `README.md`: add SO-101 to the Supported Hardware line; repoint the SoTA Models table's links from orphan stubs to the richer existing tutorial pages.
-- Move `envhub_isaaclab_arena.mdx` into Simulation (one-line YAML change).
-- Add 2-4 orientation sentences + links to the top of `index.mdx` (installation, cheat-sheet, "I have hardware" / "I don't" / "I want to contribute") without touching its marketing framing.
-- Publish a single "Choosing a policy" page that ports the already-written decision rules from `AGENT_GUIDE.md` §6 — highest-leverage fix available.
-- Start a `docs/source/_redirects.yml` now, before the orphan cleanup above creates the first real dead links.
-- No action needed on the `contributing.md` "duplication" claim — verified it's a working symlink, not an anti-pattern.
-
-### 2. Structural / UX changes (need a proposal + maintainer buy-in, phase as separate PRs)
-- Split "Tutorials" into an explicit beginner-facing section vs. an "Advanced & Research/Contributor" section (mechanical YAML reorg, no content rewrites).
-- Add a "Teleoperators" index page listing all types with one-line descriptions and links to wherever each is actually documented today.
-- Re-group the flat Policies/Benchmarks/Robots lists into sub-headings reusing the taxonomy `README.md` already has.
-- Treat a dedicated "Motors" section, deeper sub-grouping, and closing the remaining teleoperator/robot coverage gaps as a phased backlog of independently reviewable PRs rather than one restructure.
-- Standardize SO-101/SO101 naming and (after confirming actual doc-builder rendering behavior) the admonition syntax, each as one mechanical, low-risk PR.
-- Longer-term and biggest-ticket: port more of `AGENT_GUIDE.md`'s procedural content (training duration heuristics, eval targets, data-collection tips) into the published site.
-
-### 3. Maintainability / process changes (prevent debt from reaccumulating at the current growth rate)
-- Port `adding_benchmarks.mdx`'s explicit "writing a doc page" checklist/template into `bring_your_own_policies.mdx` and `integrate_hardware.mdx`.
-- Add a CODEOWNERS entry for `docs/source/` (and ideally `_toctree.yml` specifically) so docs PRs get routed to a real reviewer.
-- Link the extension guides from `CONTRIBUTING.md` and state a docs-required policy there.
-- Add a lightweight CI/pre-commit script (not a full doc-builder run) that fails when (a) a `docs/source/*.mdx` file isn't reachable from `_toctree.yml`, or (b) a new `register_subclass` policy/robot/teleoperator/env lands with no corresponding doc file in the same diff.
-- Decide the fate of the `policy_*_README.md` symlink convention going forward: it is currently self-perpetuating because `bring_your_own_policies.mdx`'s own checklist instructs new contributors to create the file that ends up orphaned. Either fold citation/paper content into the main `.mdx` tutorial, or wire the stub into the toctree as a linked citation anchor.
-- Make the PR template's "Documentation updated" checkbox actionable (e.g., "(N/A if this PR only touches tests/CI/refactors)").
-- Defer heavier generated-registry/support-matrix tooling (Transformers/Ultralytics-style single source of truth) until closer to 1.0.
-
----
-
-## Notes on cross-checking
-
-Five independent audit passes fed this report; two disagreements were resolved during synthesis:
-- **Toctree section count** — 15 is correct (independently parsed twice from `_toctree.yml`).
-- **`contributing.md`** — it's a working symlink, not a hand-duplicated anti-pattern; one pass's `diff`-based claim didn't survive checking `find -type l`.
-- **Orphan-README "sync mechanism"** — mostly fixed at the `src/`↔`docs` layer via symlinks, but that just moved the unresolved drift to the docs-side stub's absence from the toctree, and left old pre-symlink stubs (MolmoAct2) as dead leftovers.
-- **Teleoperators coverage** — the nav-section framing ("~2/16") and the "mentioned somewhere" framing are both true; the real gap is discoverability, not content.

docs/source/act.mdx

@@ -76,7 +76,7 @@ If your local computer doesn't have a powerful GPU, you can utilize Google Colab
 
 ## Evaluating ACT
 
-Once training is complete, you can evaluate your ACT policy using the `lerobot-rollout` command with your trained policy. This will run inference and record evaluation episodes:
+Once training is complete, you can evaluate your ACT policy using the `lerobot-record` command with your trained policy. This will run inference and record evaluation episodes:
 
 ```bash
 lerobot-rollout \

docs/source/earthrover_mini_plus.mdx

@@ -194,8 +194,8 @@ lerobot-record \
     --dataset.single_task="Navigate around obstacles" \
     --dataset.streaming_encoding=true \
     --dataset.encoder_threads=2 \
+    # --dataset.rgb_encoder.vcodec=auto \
     --display_data=true
-    # Optionally, set --dataset.rgb_encoder.vcodec=auto to pick a specific video codec
 ```
 
 Replace `your_username/dataset_name` with your Hugging Face username and a name for your dataset.

docs/source/hope_jr.mdx

@@ -232,8 +232,8 @@ lerobot-record \
     --dataset.private=true \
     --dataset.streaming_encoding=true \
     --dataset.encoder_threads=2 \
+    # --dataset.rgb_encoder.vcodec=auto \
     --display_data=true
-    # Optionally, set --dataset.rgb_encoder.vcodec=auto to pick a specific video codec
 ```
 
 ### Replay
@@ -278,6 +278,6 @@ lerobot-record \
   --dataset.num_episodes=10 \
   --dataset.streaming_encoding=true \
   --dataset.encoder_threads=2 \
+  # --dataset.rgb_encoder.vcodec=auto \
   --policy.path=outputs/train/hopejr_hand/checkpoints/last/pretrained_model
-  # Optionally, set --dataset.rgb_encoder.vcodec=auto to pick a specific video codec
 ```

docs/source/il_robots.mdx

@@ -207,8 +207,8 @@ lerobot-record \
     --dataset.num_episodes=5 \
     --dataset.single_task="Grab the black cube" \
     --dataset.streaming_encoding=true \
+    # --dataset.rgb_encoder.vcodec=auto \
     --dataset.encoder_threads=2
-    # Optionally, set --dataset.rgb_encoder.vcodec=auto to pick a specific video codec
 ```
 </hfoption>
 <hfoption id="API example">
@@ -418,7 +418,7 @@ If you want to dive deeper into this important topic, you can check out the [blo
 
 ## Visualize a dataset
 
-If you uploaded your dataset to the hub with `--dataset.push_to_hub=true`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
+If you uploaded your dataset to the hub with `--control.push_to_hub=true`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
 
 ```bash
 echo ${HF_USER}/so101_test

docs/source/lerobot-dataset-v3.mdx

@@ -44,8 +44,8 @@ lerobot-record \
   --dataset.num_episodes=5 \
   --dataset.single_task="Grab the black cube" \
   --dataset.streaming_encoding=true \
+  # --dataset.rgb_encoder.vcodec=auto \
   --dataset.encoder_threads=2
-  # Optionally, set --dataset.rgb_encoder.vcodec=auto to pick a specific video codec
 ```
 
 See the [recording guide](./il_robots#record-a-dataset) for more details.

docs/source/openarm.mdx

@@ -205,10 +205,9 @@ lerobot-record \
     --teleop.type=openarm_leader \
     --teleop.port=can1 \
     --teleop.id=my_leader \
-    --dataset.repo_id=my_hf_username/my_openarm_dataset \
+    --repo-id=my_hf_username/my_openarm_dataset \
-    --dataset.single_task="Grab the black cube" \
+    --fps=30 \
-    --dataset.fps=30 \
+    --num-episodes=10
-    --dataset.num_episodes=10
 ```
 
 ## Configuration Options

docs/source/reachy2.mdx

@@ -161,8 +161,8 @@ lerobot-record \
     --dataset.private=true \
     --dataset.streaming_encoding=true \
     --dataset.encoder_threads=2 \
+    # --dataset.rgb_encoder.vcodec=auto \
     --display_data=true
-    # Optionally, set --dataset.rgb_encoder.vcodec=auto to pick a specific video codec
 ```
 
 #### Specific Options
@@ -203,8 +203,8 @@ lerobot-record \
     --dataset.private=true \
     --dataset.streaming_encoding=true \
     --dataset.encoder_threads=2 \
+    # --dataset.rgb_encoder.vcodec=auto \
     --display_data=true
-    # Optionally, set --dataset.rgb_encoder.vcodec=auto to pick a specific video codec
 ```
 
 ##### `--robot.use_external_commands`

docs/source/smolvla.mdx

@@ -100,19 +100,20 @@ Once you are logged in, you can run inference in your setup by doing:
 lerobot-rollout \
   --strategy.type=base \
   --robot.type=so101_follower \
-  --robot.port=/dev/ttyACM0 \
+  --robot.port=/dev/ttyACM0 \ # <- Use your port
-  --robot.id=my_blue_follower_arm \
+  --robot.id=my_blue_follower_arm \ # <- Use your robot id
-  --robot.cameras="{ front: {type: opencv, index_or_path: 8, width: 640, height: 480, fps: 30}}" \
+  --robot.cameras="{ front: {type: opencv, index_or_path: 8, width: 640, height: 480, fps: 30}}" \ # <- Use your cameras
-  --task="Grasp a lego block and put it in the bin." \
+  --task="Grasp a lego block and put it in the bin." \ # <- Use the same task description you used in your dataset recording
-  --policy.path=HF_USER/FINETUNE_MODEL_NAME
+  # <- RTC optional, use when running on low power hardware \
+  # --inference.type=rtc \
+  # --inference.rtc.execution_horizon=10 \
+  # --inference.rtc.max_guidance_weight=10.0 \
+  # <- Teleop optional if you want to teleoperate in between episodes \
+  # --teleop.type=so100_leader \
+  # --teleop.port=/dev/ttyACM0 \
+  # --teleop.id=my_red_leader_arm \
+  # --display_data=true #optional use if you want to see the camera stream \
+  --policy.path=HF_USER/FINETUNE_MODEL_NAME # <- Use your fine-tuned model
 ```
 
-Replace `--robot.port`, `--robot.id`, `--robot.cameras`, `--task`, and `--policy.path` with your own port, robot ID, camera setup, task description (matching what you used when recording your dataset), and fine-tuned model repo ID.
-
-A few optional flags you can add to the command above:
-
-- **RTC** (useful on low-power hardware): `--inference.type=rtc --inference.rtc.execution_horizon=10 --inference.rtc.max_guidance_weight=10.0`
-- **Teleoperate in between episodes**: `--teleop.type=so100_leader --teleop.port=/dev/ttyACM0 --teleop.id=my_red_leader_arm`
-- **See the camera stream**: `--display_data=true`
-
 Depending on your evaluation setup, you can configure the duration and the number of episodes to record for your evaluation suite.

src/lerobot/configs/train.py

@@ -14,7 +14,6 @@
 import builtins
 import datetime as dt
 import json
-import multiprocessing
 import os
 import tempfile
 from dataclasses import dataclass, field
@@ -102,12 +101,6 @@ class TrainPipelineConfig(HubMixin):
     batch_size: int = 8
     prefetch_factor: int = 4
     persistent_workers: bool = True
-    # DataLoader worker start method. "spawn" is safer than "fork" with
-    # non-fork-safe libs (PyAV / torchcodec / ffmpeg), but adds some
-    # worker-startup time per run since workers re-import modules instead
-    # of inheriting parent state. Override with `--dataloader_multiprocessing_context=fork`
-    # when appropriate, or set it to `null` to use Python's platform default.
-    dataloader_multiprocessing_context: str | None = "spawn"
     steps: int = 100_000
     # Run policy in the simulation environment every N steps to measure reward/success (0 = disabled).
     env_eval_freq: int = 20_000
@@ -219,17 +212,6 @@ class TrainPipelineConfig(HubMixin):
             self.reward_model.pretrained_path = str(policy_dir)
 
     def validate(self) -> None:
-        available_contexts = multiprocessing.get_all_start_methods()
-        if (
-            self.dataloader_multiprocessing_context is not None
-            and self.dataloader_multiprocessing_context not in available_contexts
-        ):
-            raise ValueError(
-                "`dataloader_multiprocessing_context` must be None or one of "
-                f"{available_contexts} on this platform, got "
-                f"{self.dataloader_multiprocessing_context!r}."
-            )
-
         self._resolve_pretrained_from_cli()
 
         if self.policy is None and self.reward_model is None:

src/lerobot/datasets/compute_stats.py

@@ -519,6 +519,13 @@ def compute_episode_stats(
         if features[key]["dtype"] in {"string", "language"}:
             continue
 
+        # Features with a zero-width dimension contain no statistics-bearing
+        # values. Skip them like strings instead of letting
+        # get_feature_stats -> RunningQuantileStats.update reshape a size-0 array,
+        # which raises "ValueError: cannot reshape array of size 0".
+        if any(dim == 0 for dim in features[key].get("shape", ())):
+            continue
+
         if features[key]["dtype"] in ["image", "video"]:
             ep_ft_array = sample_images(data)
             axes_to_reduce = (0, 2, 3)

src/lerobot/datasets/feature_utils.py

@@ -64,12 +64,20 @@ def get_hf_features_from_features(features: dict) -> datasets.Features:
             continue
         elif ft["dtype"] == "image":
             hf_features[key] = datasets.Image()
+        elif len(ft["shape"]) > 1 and any(dim == 0 for dim in ft["shape"]):
+            raise ValueError(
+                f"Multidimensional features with a zero-width dimension are not supported: "
+                f"'{key}' has shape {ft['shape']}. Only the one-dimensional shape (0,) is supported."
+            )
         elif ft["shape"] == (1,):
             hf_features[key] = datasets.Value(dtype=ft["dtype"])
         elif len(ft["shape"]) == 1:
-            hf_features[key] = datasets.Sequence(
+            # A zero-width feature (shape=(0,)) has no fixed-size Arrow representation:
-                length=ft["shape"][0], feature=datasets.Value(dtype=ft["dtype"])
+            # pyarrow rejects a fixed-size list of length 0 ("list_size needs to be a
-            )
+            # strict positive integer"). Store it as a variable-length sequence
+            # (length=-1) so each per-frame value is simply an empty list.
+            seq_length = ft["shape"][0] if ft["shape"][0] > 0 else -1
+            hf_features[key] = datasets.Sequence(length=seq_length, feature=datasets.Value(dtype=ft["dtype"]))
         elif len(ft["shape"]) == 2:
             hf_features[key] = datasets.Array2D(shape=ft["shape"], dtype=ft["dtype"])
         elif len(ft["shape"]) == 3:

src/lerobot/envs/metaworld.py

@@ -155,7 +155,6 @@ class MetaworldEnv(gym.Env):
             env.model.cam_pos[2] = [0.75, 0.075, 0.7]
         env.reset()
         env._freeze_rand_vec = False  # otherwise no randomization
-        env.seeded_rand_vec = True  # use seeded RNG so reset(seed=X) controls object positions
         self._env = env
 
     def render(self) -> np.ndarray:
@@ -221,8 +220,6 @@ class MetaworldEnv(gym.Env):
         self._ensure_env()
         super().reset(seed=seed)
 
-        if seed is not None:
-            self._env.seed(seed)
         raw_obs, info = self._env.reset(seed=seed)
 
         observation = self._format_raw_obs(raw_obs)

src/lerobot/policies/factory.py

@@ -177,7 +177,6 @@ def make_pre_post_processors(
             return make_groot_pre_post_processors_from_pretrained(
                 config=policy_cfg,
                 pretrained_path=pretrained_path,
-                revision=pretrained_revision,
                 dataset_stats=kwargs.get("dataset_stats"),
                 dataset_meta=kwargs.get("dataset_meta"),
                 preprocessor_overrides=kwargs.get("preprocessor_overrides"),

src/lerobot/policies/groot/processor_groot.py

@@ -475,7 +475,6 @@ def make_groot_pre_post_processors_from_pretrained(
     config: GrootConfig,
     pretrained_path: str,
     *,
-    revision: str | None = None,
     dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
     dataset_meta: Any | None = None,
     preprocessor_overrides: dict[str, Any] | None = None,
@@ -512,7 +511,6 @@ def make_groot_pre_post_processors_from_pretrained(
 
     preprocessor, postprocessor = _load_groot_processor_pipelines(
         pretrained_path,
-        revision=revision,
         preprocessor_overrides=preprocessor_overrides,
         postprocessor_overrides=postprocessor_overrides,
         preprocessor_config_filename=preprocessor_config_filename,
@@ -528,7 +526,6 @@ def make_groot_pre_post_processors_from_pretrained(
 def _load_groot_processor_pipelines(
     pretrained_path: str,
     *,
-    revision: str | None,
     preprocessor_overrides: dict[str, Any],
     postprocessor_overrides: dict[str, Any],
     preprocessor_config_filename: str,
@@ -543,7 +540,6 @@ def _load_groot_processor_pipelines(
     preprocessor = PolicyProcessorPipeline.from_pretrained(
         pretrained_model_name_or_path=pretrained_path,
         config_filename=preprocessor_config_filename,
-        revision=revision,
         overrides=preprocessor_overrides,
         to_transition=batch_to_transition,
         to_output=transition_to_batch,
@@ -551,7 +547,6 @@ def _load_groot_processor_pipelines(
     postprocessor = PolicyProcessorPipeline.from_pretrained(
         pretrained_model_name_or_path=pretrained_path,
         config_filename=postprocessor_config_filename,
-        revision=revision,
         overrides=postprocessor_overrides,
         to_transition=policy_action_to_transition,
         to_output=transition_to_policy_action,

src/lerobot/processor/pipeline.py

@@ -713,8 +713,6 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
             ProcessorMigrationError: If the model requires migration to processor format.
         """
         model_id = str(pretrained_model_name_or_path)
-        model_path = Path(model_id)
-        is_local_source = model_path.is_dir() or model_path.is_file()
         hub_download_kwargs = {
             "force_download": force_download,
             "resume_download": resume_download,
@@ -733,7 +731,7 @@ 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, hub_download_kwargs
         )
 
         # 4. Validate that all overrides were used
@@ -923,7 +921,6 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
         model_id: str,
         base_path: Path | None,
         hub_download_kwargs: dict[str, Any],
-        is_local_source: bool = False,
     ) -> tuple[list[ProcessorStep], set[str]]:
         """Build all processor steps with overrides and state loading.
 
@@ -947,7 +944,7 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
         3. **State Loading** (via _load_step_state):
            - **If step has "state_file"**: Load tensor state from .safetensors
            - **Local first**: Check base_path/state_file.safetensors
-           - **Hub fallback**: Download state file if the pipeline was loaded from the Hub
+           - **Hub fallback**: Download state file if not found locally
            - **Optional**: Only load if step has load_state_dict method
 
         4. **Override Tracking**:
@@ -965,7 +962,6 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
             model_id: The model identifier (needed for Hub state file downloads)
             base_path: Local directory path for finding state files
             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.
 
         Returns:
             Tuple of (instantiated_steps_list, unused_override_keys)
@@ -979,9 +975,7 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
         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(
+            cls._load_step_state(step_instance, step_entry, model_id, base_path, hub_download_kwargs)
-                step_instance, step_entry, model_id, base_path, hub_download_kwargs, is_local_source
-            )
 
         return steps, remaining_override_keys
 
@@ -1145,7 +1139,6 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
         model_id: str,
         base_path: Path | None,
         hub_download_kwargs: dict[str, Any],
-        is_local_source: bool = False,
     ) -> None:
         """Load state dictionary for a processor step if available.
 
@@ -1164,7 +1157,7 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
            - **Use case**: Loading from local saved model directory
 
         2. **Hub download fallback**: Download state file from repository
-           - **When triggered**: Local file not found and the pipeline source is a Hub repo
+           - **When triggered**: Local file not found or base_path is None
            - **Process**: Use hf_hub_download with same parameters as config
            - **Example**: Download "normalize_step_0.safetensors" from "user/repo"
            - **Result**: Downloaded to local cache, path returned
@@ -1185,7 +1178,6 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
             model_id: The model identifier (used for Hub downloads if needed)
             base_path: Local directory path for finding state files (None for Hub-only)
             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.
 
         Note:
             This method modifies step_instance in-place and returns None.
@@ -1199,12 +1191,6 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
         # Try local file first
         if base_path and (base_path / state_filename).exists():
             state_path = str(base_path / state_filename)
-        elif is_local_source:
-            state_path = base_path / state_filename if base_path else Path(state_filename)
-            raise FileNotFoundError(
-                f"State file '{state_filename}' was not found for local processor pipeline "
-                f"'{model_id}' at '{state_path}'."
-            )
         else:
             # Download from Hub
             state_path = hf_hub_download(

src/lerobot/robots/lekiwi/lekiwi_client.py

@@ -323,10 +323,6 @@ class LeKiwiClient(Robot):
             np.ndarray: the action sent to the motors, potentially clipped.
         """
 
-        # Action values may be torch tensors (e.g. replayed from a dataset) or numpy
-        # scalars; json.dumps only serializes Python primitives, so coerce each value to a
-        # plain float before sending.
-        action = {key: float(value) for key, value in action.items()}
         self.zmq_cmd_socket.send_string(json.dumps(action))  # action is in motor space
 
         # TODO(Steven): Remove the np conversion when it is possible to record a non-numpy array value

src/lerobot/rollout/configs.py

@@ -326,17 +326,8 @@ class RolloutConfig:
 
         policy_path = parser.get_path_arg("policy")
         if policy_path:
-            yaml_overrides = parser.get_yaml_overrides("policy")
+            cli_overrides = parser.get_cli_overrides("policy")
-            cli_overrides = parser.get_cli_overrides("policy") or []
+            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
-            policy_overrides = yaml_overrides + cli_overrides
-            pretrained_revision = parser.parse_arg("pretrained_revision", cli_overrides)
-            if pretrained_revision is None:
-                pretrained_revision = parser.parse_arg("pretrained_revision", yaml_overrides)
-            self.policy = PreTrainedConfig.from_pretrained(
-                policy_path,
-                revision=pretrained_revision,
-                cli_overrides=policy_overrides,
-            )
             self.policy.pretrained_path = policy_path
         if self.policy is None:
             raise ValueError("--policy.path is required for rollout")

src/lerobot/rollout/context.py

@@ -27,7 +27,7 @@ from threading import Event
 
 import torch
 
-from lerobot.configs import FeatureType, PreTrainedConfig
+from lerobot.configs import FeatureType
 from lerobot.datasets import (
     LeRobotDataset,
     aggregate_pipeline_dataset_features,
@@ -159,35 +159,6 @@ class RolloutContext:
 # ---------------------------------------------------------------------------
 
 
-def _load_pretrained_policy(policy_config: PreTrainedConfig) -> PreTrainedPolicy:
-    """Load policy weights, keeping adapter and base-model revisions independent."""
-    pretrained_revision = policy_config.pretrained_revision
-    policy_class = get_policy_class(policy_config.type)
-
-    if not policy_config.use_peft:
-        return policy_class.from_pretrained(
-            policy_config.pretrained_path,
-            config=policy_config,
-            revision=pretrained_revision,
-        )
-
-    from peft import PeftConfig, PeftModel
-
-    peft_path = policy_config.pretrained_path
-    peft_config = PeftConfig.from_pretrained(peft_path, revision=pretrained_revision)
-    policy = policy_class.from_pretrained(
-        pretrained_name_or_path=peft_config.base_model_name_or_path,
-        config=policy_config,
-        revision=peft_config.revision,
-    )
-    return PeftModel.from_pretrained(
-        policy,
-        peft_path,
-        config=peft_config,
-        revision=pretrained_revision,
-    )
-
-
 def build_rollout_context(
     cfg: RolloutConfig,
     shutdown_event: Event,
@@ -205,6 +176,7 @@ def build_rollout_context(
     # --- 1. Policy (heavy I/O, but no hardware yet) -------------------
     logger.info("Loading policy from '%s'...", cfg.policy.pretrained_path)
     policy_config = cfg.policy
+    policy_class = get_policy_class(policy_config.type)
 
     if hasattr(policy_config, "compile_model"):
         policy_config.compile_model = cfg.use_torch_compile
@@ -215,7 +187,17 @@ def build_rollout_context(
             "Please use `cpu` or `cuda` backend."
         )
 
-    policy = _load_pretrained_policy(policy_config)
+    if policy_config.use_peft:
+        from peft import PeftConfig, PeftModel
+
+        peft_path = policy_config.pretrained_path
+        peft_config = PeftConfig.from_pretrained(peft_path)
+        policy = policy_class.from_pretrained(
+            pretrained_name_or_path=peft_config.base_model_name_or_path, config=policy_config
+        )
+        policy = PeftModel.from_pretrained(policy, peft_path, config=peft_config)
+    else:
+        policy = policy_class.from_pretrained(policy_config.pretrained_path, config=policy_config)
 
     if is_rtc:
         policy.config.rtc_config = cfg.inference.rtc
@@ -410,7 +392,6 @@ def build_rollout_context(
     preprocessor, postprocessor = make_pre_post_processors(
         policy_cfg=policy_config,
         pretrained_path=cfg.policy.pretrained_path,
-        pretrained_revision=policy_config.pretrained_revision,
         dataset_stats=dataset_stats,
         preprocessor_overrides={
             "device_processor": {"device": cfg.device},

src/lerobot/scripts/lerobot_replay.py

@@ -61,7 +61,6 @@ from lerobot.robots import (  # noqa: F401
     earthrover_mini_plus,
     hope_jr,
     koch_follower,
-    lekiwi,
     make_robot_from_config,
     omx_follower,
     openarm_follower,

src/lerobot/scripts/lerobot_train.py

@@ -71,16 +71,6 @@ from lerobot.utils.utils import (
 from .lerobot_eval import eval_policy_all
 
 
-def _dataloader_worker_kwargs(cfg: TrainPipelineConfig) -> dict[str, Any]:
-    """Return worker-only DataLoader options, disabling them for single-process loading."""
-    workers_enabled = cfg.num_workers > 0
-    return {
-        "prefetch_factor": cfg.prefetch_factor if workers_enabled else None,
-        "persistent_workers": cfg.persistent_workers and workers_enabled,
-        "multiprocessing_context": cfg.dataloader_multiprocessing_context if workers_enabled else None,
-    }
-
-
 def update_policy(
     train_metrics: MetricsTracker,
     policy: PreTrainedPolicy,
@@ -483,7 +473,8 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
         pin_memory=device.type == "cuda",
         drop_last=False,
         collate_fn=collate_fn,
-        **_dataloader_worker_kwargs(cfg),
+        prefetch_factor=cfg.prefetch_factor if cfg.num_workers > 0 else None,
+        persistent_workers=cfg.persistent_workers and cfg.num_workers > 0,
     )
 
     # Build eval dataloader if a held-out split exists
@@ -509,7 +500,8 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
             pin_memory=device.type == "cuda",
             drop_last=False,
             collate_fn=eval_collate_fn,
-            **_dataloader_worker_kwargs(cfg),
+            prefetch_factor=cfg.prefetch_factor if cfg.num_workers > 0 else None,
+            persistent_workers=cfg.persistent_workers and cfg.num_workers > 0,
         )
 
     # Prepare everything with accelerator

tests/datasets/test_compute_stats.py

@@ -687,6 +687,26 @@ def test_compute_episode_stats_string_features_skipped():
     assert "q01" in stats["action"]
 
 
+@pytest.mark.parametrize("shape", [(0,), (0, 2), (2, 0), (1, 0, 2)])
+def test_compute_episode_stats_zero_width_feature_skipped(shape):
+    """Features with any zero-width dimension carry no values and are skipped."""
+    episode_data = {
+        "action": np.random.normal(0, 1, (100, 5)).astype(np.float32),
+        "target": np.zeros((100, *shape), dtype=np.float32),
+    }
+    features = {
+        "action": {"dtype": "float32", "shape": (5,)},
+        "target": {"dtype": "float32", "shape": shape},
+    }
+
+    stats = compute_episode_stats(episode_data, features)
+
+    # Zero-width features are skipped, just like strings; non-empty features are unaffected.
+    assert "target" not in stats
+    assert "action" in stats
+    assert "q01" in stats["action"]
+
+
 def test_aggregate_feature_stats_with_quantiles():
     """Test aggregating feature stats that include quantiles."""
     stats_ft_list = [

tests/datasets/test_dataset_writer.py

@@ -27,6 +27,7 @@ pytest.importorskip("datasets", reason="datasets is required (install lerobot[da
 
 from lerobot.configs import VideoEncoderConfig
 from lerobot.datasets.dataset_writer import _encode_video_worker
+from lerobot.datasets.feature_utils import get_hf_features_from_features
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.utils import DEFAULT_IMAGE_PATH
 from tests.fixtures.constants import DEFAULT_FPS, DUMMY_REPO_ID
@@ -189,6 +190,36 @@ def test_save_multiple_episodes(tmp_path):
     assert dataset.meta.total_frames == total_frames
 
 
+def test_save_episode_with_zero_width_feature(tmp_path):
+    """A one-dimensional empty numeric feature round-trips and has no statistics."""
+    features = {
+        **SIMPLE_FEATURES,
+        "target": {"dtype": "float32", "shape": (0,), "names": None},
+    }
+    root = tmp_path / "ds"
+    dataset = LeRobotDataset.create(repo_id=DUMMY_REPO_ID, fps=DEFAULT_FPS, features=features, root=root)
+    for _ in range(4):
+        dataset.add_frame(_make_frame(features))
+    dataset.save_episode()
+    dataset.finalize()
+
+    assert dataset.meta.total_episodes == 1
+    assert dataset.meta.total_frames == 4
+
+    reloaded = LeRobotDataset(repo_id=DUMMY_REPO_ID, root=root)
+    target = np.asarray(reloaded[0]["target"])
+    assert target.shape == (0,)
+    assert "target" not in (reloaded.meta.stats or {})
+
+
+@pytest.mark.parametrize("shape", [(0, 2), (2, 0), (1, 0, 2)])
+def test_multidimensional_zero_width_feature_rejected(shape):
+    features = {"target": {"dtype": "float32", "shape": shape, "names": None}}
+
+    with pytest.raises(ValueError, match="Multidimensional features with a zero-width dimension"):
+        get_hf_features_from_features(features)
+
+
 # ── clear / lifecycle ────────────────────────────────────────────────
 
 

tests/processor/test_pipeline_from_pretrained_helpers.py

@@ -26,17 +26,8 @@ import tempfile
 from pathlib import Path
 
 import pytest
-import torch
-from safetensors.torch import save_file
 
-from lerobot.configs import PipelineFeatureType, PolicyFeature
+from lerobot.processor.pipeline import DataProcessorPipeline, ProcessorMigrationError
-from lerobot.processor.pipeline import (
-    DataProcessorPipeline,
-    ProcessorMigrationError,
-    ProcessorStep,
-    ProcessorStepRegistry,
-)
-from lerobot.types import EnvTransition
 
 # Simplified Config Loading Tests
 
@@ -107,140 +98,6 @@ def test_load_config_nonexistent_path_tries_hub():
         DataProcessorPipeline._load_config("nonexistent/path", "processor.json", {})
 
 
-def test_from_pretrained_local_directory_missing_state_does_not_call_hub(monkeypatch):
-    """Local processor dirs must fail locally when a state file is missing."""
-
-    @ProcessorStepRegistry.register("local_missing_state_step")
-    class LocalMissingStateStep(ProcessorStep):
-        def __call__(self, transition: EnvTransition) -> EnvTransition:
-            return transition
-
-        def transform_features(
-            self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
-        ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
-            return features
-
-        def load_state_dict(self, state: dict[str, torch.Tensor]) -> None:
-            pass
-
-    try:
-        with tempfile.TemporaryDirectory() as tmp_dir:
-            tmp_path = Path(tmp_dir)
-            config = {
-                "name": "LocalMissingStatePipeline",
-                "steps": [{"registry_name": "local_missing_state_step", "state_file": "missing.safetensors"}],
-            }
-            (tmp_path / "processor.json").write_text(json.dumps(config))
-
-            def fail_hub_download(*args, **kwargs):
-                pytest.fail("local missing processor state should not call hf_hub_download")
-
-            monkeypatch.setattr("lerobot.processor.pipeline.hf_hub_download", fail_hub_download)
-
-            with pytest.raises(FileNotFoundError, match="missing.safetensors.*local processor pipeline"):
-                DataProcessorPipeline.from_pretrained(tmp_path, config_filename="processor.json")
-    finally:
-        ProcessorStepRegistry.unregister("local_missing_state_step")
-
-
-def test_from_pretrained_local_config_file_missing_state_does_not_call_hub(monkeypatch):
-    """Local single-file processor configs must also keep missing state resolution local."""
-
-    @ProcessorStepRegistry.register("local_file_missing_state_step")
-    class LocalFileMissingStateStep(ProcessorStep):
-        def __call__(self, transition: EnvTransition) -> EnvTransition:
-            return transition
-
-        def transform_features(
-            self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
-        ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
-            return features
-
-        def load_state_dict(self, state: dict[str, torch.Tensor]) -> None:
-            pass
-
-    try:
-        with tempfile.TemporaryDirectory() as tmp_dir:
-            tmp_path = Path(tmp_dir)
-            config_path = tmp_path / "processor.json"
-            config = {
-                "name": "LocalFileMissingStatePipeline",
-                "steps": [
-                    {"registry_name": "local_file_missing_state_step", "state_file": "missing.safetensors"}
-                ],
-            }
-            config_path.write_text(json.dumps(config))
-
-            def fail_hub_download(*args, **kwargs):
-                pytest.fail("local missing processor state should not call hf_hub_download")
-
-            monkeypatch.setattr("lerobot.processor.pipeline.hf_hub_download", fail_hub_download)
-
-            with pytest.raises(FileNotFoundError, match="missing.safetensors.*local processor pipeline"):
-                DataProcessorPipeline.from_pretrained(config_path, config_filename="ignored.json")
-    finally:
-        ProcessorStepRegistry.unregister("local_file_missing_state_step")
-
-
-def test_from_pretrained_hub_source_missing_local_state_still_calls_hub(monkeypatch, tmp_path):
-    """Hub sources still fall back to hf_hub_download for state files."""
-
-    @ProcessorStepRegistry.register("hub_state_step")
-    class HubStateStep(ProcessorStep):
-        def __init__(self):
-            self.value = torch.tensor(0)
-
-        def __call__(self, transition: EnvTransition) -> EnvTransition:
-            return transition
-
-        def transform_features(
-            self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
-        ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
-            return features
-
-        def load_state_dict(self, state: dict[str, torch.Tensor]) -> None:
-            self.value = state["value"]
-
-    try:
-        state_path = tmp_path / "downloaded.safetensors"
-        save_file({"value": torch.tensor(7)}, state_path)
-        loaded_config = {
-            "name": "HubStatePipeline",
-            "steps": [{"registry_name": "hub_state_step", "state_file": "hub_state.safetensors"}],
-        }
-        calls = []
-
-        def fake_load_config(cls, model_id, config_filename, hub_download_kwargs):
-            return loaded_config, tmp_path / "hub_cache"
-
-        def fake_hub_download(**kwargs):
-            calls.append(kwargs)
-            return str(state_path)
-
-        monkeypatch.setattr(DataProcessorPipeline, "_load_config", classmethod(fake_load_config))
-        monkeypatch.setattr("lerobot.processor.pipeline.hf_hub_download", fake_hub_download)
-
-        pipeline = DataProcessorPipeline.from_pretrained("user/repo", config_filename="processor.json")
-
-        assert calls == [
-            {
-                "repo_id": "user/repo",
-                "filename": "hub_state.safetensors",
-                "repo_type": "model",
-                "force_download": False,
-                "resume_download": None,
-                "proxies": None,
-                "token": None,
-                "cache_dir": None,
-                "local_files_only": False,
-                "revision": None,
-            }
-        ]
-        assert pipeline.steps[0].value.item() == 7
-    finally:
-        ProcessorStepRegistry.unregister("hub_state_step")
-
-
 # Config Validation Tests
 
 

tests/test_rollout.py

@@ -17,8 +17,6 @@
 from __future__ import annotations
 
 import dataclasses
-import sys
-from types import SimpleNamespace
 from unittest.mock import MagicMock
 
 import pytest
@@ -108,109 +106,6 @@ def test_sentry_config_defaults():
     assert cfg.target_video_file_size_mb is None
 
 
-def test_rollout_config_passes_policy_pretrained_revision(monkeypatch):
-    from lerobot.configs import PreTrainedConfig, parser
-    from lerobot.rollout import RolloutConfig
-    from tests.mocks.mock_robot import MockRobotConfig
-
-    captured = {}
-
-    def fake_from_pretrained(cls, pretrained_name_or_path, **kwargs):
-        captured["pretrained_name_or_path"] = pretrained_name_or_path
-        captured.update(kwargs)
-        return SimpleNamespace(device="cpu", pretrained_revision=kwargs["revision"])
-
-    monkeypatch.setattr(parser, "get_yaml_overrides", lambda _: ["--pretrained_revision=yaml-sha"])
-    monkeypatch.setattr(
-        sys,
-        "argv",
-        ["lerobot-rollout", "--policy.path=user/policy", "--policy.pretrained_revision=cli-sha"],
-    )
-    monkeypatch.setattr(PreTrainedConfig, "from_pretrained", classmethod(fake_from_pretrained))
-
-    cfg = RolloutConfig(robot=MockRobotConfig())
-
-    assert captured["pretrained_name_or_path"] == "user/policy"
-    assert captured["revision"] == "cli-sha"
-    assert captured["cli_overrides"] == [
-        "--pretrained_revision=yaml-sha",
-        "--pretrained_revision=cli-sha",
-    ]
-    assert cfg.policy.pretrained_path == "user/policy"
-    assert cfg.policy.pretrained_revision == "cli-sha"
-
-
-def test_load_pretrained_policy_passes_revision(monkeypatch):
-    import lerobot.rollout.context as rollout_context
-
-    policy_config = SimpleNamespace(
-        type="mock",
-        use_peft=False,
-        pretrained_path="user/policy",
-        pretrained_revision="policy-sha",
-    )
-    policy_class = MagicMock()
-    loaded_policy = MagicMock()
-    policy_class.from_pretrained.return_value = loaded_policy
-    monkeypatch.setattr(rollout_context, "get_policy_class", lambda _: policy_class)
-
-    policy = rollout_context._load_pretrained_policy(policy_config)
-
-    assert policy is loaded_policy
-    policy_class.from_pretrained.assert_called_once_with(
-        "user/policy",
-        config=policy_config,
-        revision="policy-sha",
-    )
-
-
-def test_load_pretrained_peft_policy_keeps_adapter_and_base_revisions_separate(monkeypatch):
-    import lerobot.rollout.context as rollout_context
-
-    policy_config = SimpleNamespace(
-        type="mock",
-        use_peft=True,
-        pretrained_path="user/adapter",
-        pretrained_revision="adapter-sha",
-    )
-    policy_class = MagicMock()
-    base_policy = MagicMock()
-    policy_class.from_pretrained.return_value = base_policy
-    monkeypatch.setattr(rollout_context, "get_policy_class", lambda _: policy_class)
-
-    peft_config = SimpleNamespace(
-        base_model_name_or_path="user/base-policy",
-        revision="base-sha",
-    )
-    peft_config_from_pretrained = MagicMock(return_value=peft_config)
-    adapted_policy = MagicMock()
-    peft_model_from_pretrained = MagicMock(return_value=adapted_policy)
-    monkeypatch.setitem(
-        sys.modules,
-        "peft",
-        SimpleNamespace(
-            PeftConfig=SimpleNamespace(from_pretrained=peft_config_from_pretrained),
-            PeftModel=SimpleNamespace(from_pretrained=peft_model_from_pretrained),
-        ),
-    )
-
-    policy = rollout_context._load_pretrained_policy(policy_config)
-
-    assert policy is adapted_policy
-    peft_config_from_pretrained.assert_called_once_with("user/adapter", revision="adapter-sha")
-    policy_class.from_pretrained.assert_called_once_with(
-        pretrained_name_or_path="user/base-policy",
-        config=policy_config,
-        revision="base-sha",
-    )
-    peft_model_from_pretrained.assert_called_once_with(
-        base_policy,
-        "user/adapter",
-        config=peft_config,
-        revision="adapter-sha",
-    )
-
-
 # ---------------------------------------------------------------------------
 # RolloutRingBuffer
 # ---------------------------------------------------------------------------