adding relative joint exclude test

* refactor(processors): share policy pipeline builders

* Apply suggestions from code review

Co-authored-by: Martino Russi <77496684+nepyope@users.noreply.github.com>
Signed-off-by: Steven Palma <imstevenpmwork@ieee.org>

* fix(processor): solve style after commit suggestions

---------

Signed-off-by: Steven Palma <imstevenpmwork@ieee.org>
Co-authored-by: Martino Russi <77496684+nepyope@users.noreply.github.com>

docs/source/annotation_pipeline.mdx

@@ -81,6 +81,12 @@ merged. Both prompts also carry a causal **event-boundary** definition (a
 new event starts when an object becomes held / is released / reaches a new
 location / a lid changes state / contents move) to sharpen where cuts land.
 
+Optionally, a third **seeded-relabel** pass (`--plan.subtask_seeded_relabel`)
+revisits each span with its previous/current/next segment contact sheets and
+minimally corrects the label, using the first label as a prior — it keeps the
+boundaries fixed and only sharpens wording, at the cost of one extra call per
+subtask.
+
 The resulting spans are then stitched into a gap-free, full-episode
 cover, so **every frame has exactly one active subtask**. See
 [`run_hf_job.py`](https://github.com/huggingface/lerobot/blob/main/examples/annotations/run_hf_job.py)
@@ -157,30 +163,33 @@ Every module is on by default and can be toggled independently (set to
 
 ### The VLM (`--vlm.*`)
 
-| Flag                       | Default            | What it does                                                                        |
-| -------------------------- | ------------------ | ----------------------------------------------------------------------------------- |
-| `--vlm.model_id`           | `Qwen/Qwen3.6-27B` | The model to serve and prompt.                                                      |
-| `--vlm.camera_key`         | first `images.*`   | Which camera every prompt is grounded on.                                           |
-| `--vlm.serve_command`      | auto               | The exact `vllm serve …` command (set TP size, GPU memory, `--max-model-len` here). |
-| `--vlm.parallel_servers`   | `1`                | Independent servers for round-robin routing (one per GPU).                          |
-| `--vlm.num_gpus`           | `0`                | GPUs per server (`0` = one each).                                                   |
-| `--vlm.client_concurrency` | `16`               | In-flight requests across all servers.                                              |
-| `--vlm.max_new_tokens`     | `512`              | Generation cap per call.                                                            |
-| `--vlm.temperature`        | `0.2`              | Sampling temperature.                                                               |
+| Flag                       | Default            | What it does                                                                         |
+| -------------------------- | ------------------ | ------------------------------------------------------------------------------------ |
+| `--vlm.model_id`           | `Qwen/Qwen3.6-27B` | The model to serve and prompt.                                                       |
+| `--vlm.camera_key`         | first `images.*`   | Which camera every prompt is grounded on.                                            |
+| `--vlm.serve_command`      | auto               | The exact `vllm serve …` command (set TP size, GPU memory, `--max-model-len` here).  |
+| `--vlm.parallel_servers`   | `1`                | Independent servers for round-robin routing (one per GPU).                           |
+| `--vlm.num_gpus`           | `0`                | GPUs per server (`0` = one each).                                                    |
+| `--vlm.client_concurrency` | `16`               | In-flight requests across all servers.                                               |
+| `--vlm.max_new_tokens`     | `512`              | Generation cap per call.                                                             |
+| `--vlm.temperature`        | `0.2`              | Sampling temperature.                                                                |
+| `--vlm.reasoning_effort`   | `null`             | Thinking-budget hint (`low`/`medium`/`high`) forwarded to OpenAI-compatible servers. |
 
 ### Subtasks / plan / memory (`--plan.*`)
 
-| Flag                            | Default    | What it does                                                                                                              |
-| ------------------------------- | ---------- | ------------------------------------------------------------------------------------------------------------------------- |
-| `--plan.frames_per_second`      | `2.0`      | Frame sampling rate for the contact sheets (`2.0` = one frame every 0.5s).                                                |
-| `--plan.max_frames_per_prompt`  | `60`       | Frame budget per VLM call. Episodes whose sampling exceeds this are auto-windowed at the same density, then stitched.     |
-| `--plan.contact_sheet_columns`  | `5`        | Columns per contact-sheet grid (`contact_sheet_frames_per_sheet` tiles, time row-major).                                  |
-| `--plan.plan_max_steps`         | `8`        | Upper bound on subtasks per episode.                                                                                      |
-| `--plan.subtask_describe_first` | `true`     | Run the describe→segment grounding pass (best subtask quality; +1 call/episode).                                          |
-| `--plan.emit_plan`              | `true`     | Emit the numbered `plan` rows (`false` = subtasks + memory only).                                                         |
-| `--plan.emit_memory`            | `true`     | Emit the `memory` rows (`false` = subtasks + plan only); symmetric to `emit_plan`.                                        |
-| `--plan.n_task_rephrasings`     | `10`       | How many `task_aug` rephrasings to emit (`0` disables).                                                                   |
-| `--plan.derive_task_from_video` | `if_short` | Use the dataset task as-is (`off`), only when it's missing/short (`if_short`), or always re-derive from video (`always`). |
+| Flag                            | Default    | What it does                                                                                                                 |
+| ------------------------------- | ---------- | ---------------------------------------------------------------------------------------------------------------------------- |
+| `--plan.frames_per_second`      | `2.0`      | Frame sampling rate for the contact sheets (`2.0` = one frame every 0.5s).                                                   |
+| `--plan.max_frames_per_prompt`  | `60`       | Frame budget per VLM call. Episodes whose sampling exceeds this are auto-windowed at the same density, then stitched.        |
+| `--plan.contact_sheet_columns`  | `5`        | Columns per contact-sheet grid (`contact_sheet_frames_per_sheet` tiles, time row-major).                                     |
+| `--plan.plan_max_steps`         | `8`        | Upper bound on subtasks per episode.                                                                                         |
+| `--plan.subtask_describe_first` | `true`     | Run the describe→segment grounding pass (best subtask quality; +1 call/episode).                                             |
+| `--plan.subtask_seeded_relabel` | `false`    | Second pass: re-label each subtask from its prev/current/next contact sheets, seeded with the first label (+1 call/subtask). |
+| `--plan.subtask_relabel_frames` | `5`        | Frames sampled uniformly per segment sheet in the relabel pass (only used when `subtask_seeded_relabel=true`).               |
+| `--plan.emit_plan`              | `true`     | Emit the numbered `plan` rows (`false` = subtasks + memory only).                                                            |
+| `--plan.emit_memory`            | `true`     | Emit the `memory` rows (`false` = subtasks + plan only); symmetric to `emit_plan`.                                           |
+| `--plan.n_task_rephrasings`     | `10`       | How many `task_aug` rephrasings to emit (`0` disables).                                                                      |
+| `--plan.derive_task_from_video` | `if_short` | Use the dataset task as-is (`off`), only when it's missing/short (`if_short`), or always re-derive from video (`always`).    |
 
 ### Interjections + VQA
 

docs/source/bring_your_own_policies.mdx

@@ -150,14 +150,14 @@ class MyPolicy(PreTrainedPolicy):
 
 The methods called by the train/eval loops:
 
-| Method                                                            | Used by           | What it does                                                                                                                                                                                                                                         |
-| ----------------------------------------------------------------- | ----------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| `reset() -> None`                                                 | `lerobot-eval`    | Clear per-episode state at the start of each episode.                                                                                                                                                                                                |
-| `select_action(batch, **kwargs) -> Tensor`                        | `lerobot-eval`    | Return the next action `(B, action_dim)`. Called every step.                                                                                                                                                                                         |
-| `predict_action_chunk(batch, **kwargs) -> Tensor`                 | the policy itself | Return an action chunk `(B, chunk_size, action_dim)`. Currently abstract on the base class — raise `NotImplementedError` if your policy doesn't chunk.                                                                                               |
-| `forward(batch, reduction="mean") -> tuple[Tensor, dict \| None]` | `lerobot-train`   | Return `(loss, output_dict)`. Accept `reduction="none"` if you want to support per-sample weighting.                                                                                                                                                 |
-| `get_optim_params() -> dict`                                      | the optimizer     | Return `self.parameters()` for simple policies; return a named parameter dict for [multi-optimizer policies](https://github.com/huggingface/lerobot/blob/ecd38c50d7d15b4184cf42649ff1185ee2e11eeb/src/lerobot/policies/sac/modeling_sac.py#L61-L73). |
-| `update() -> None` _(optional)_                                   | `lerobot-train`   | Called after each optimizer step _if defined_. Use for EMA, target nets, replay buffers (TDMPC uses this).                                                                                                                                           |
+| Method                                                            | Used by           | What it does                                                                                                                                                                                                                                                                                 |
+| ----------------------------------------------------------------- | ----------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `reset() -> None`                                                 | `lerobot-eval`    | Clear per-episode state at the start of each episode.                                                                                                                                                                                                                                        |
+| `select_action(batch, **kwargs) -> Tensor`                        | `lerobot-eval`    | Return the next action `(B, action_dim)`. Called every step.                                                                                                                                                                                                                                 |
+| `predict_action_chunk(batch, **kwargs) -> Tensor`                 | the policy itself | Return an action chunk `(B, chunk_size, action_dim)`. Currently abstract on the base class — raise `NotImplementedError` if your policy doesn't chunk.                                                                                                                                       |
+| `forward(batch, reduction="mean") -> tuple[Tensor, dict \| None]` | `lerobot-train`   | Return `(loss, output_dict)`. Accept `reduction="none"` if you want to support per-sample weighting.                                                                                                                                                                                         |
+| `get_optim_params() -> dict`                                      | the optimizer     | Return `self.parameters()` for simple policies; return a named parameter dict for multi-optimizer policies (see `get_optim_params` in [`modeling_act.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/act/modeling_act.py) for a per-group learning-rate example). |
+| `update() -> None` _(optional)_                                   | `lerobot-train`   | Called after each optimizer step _if defined_. Use for EMA, target nets, replay buffers (TDMPC uses this).                                                                                                                                                                                   |
 
 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.
 
@@ -295,12 +295,10 @@ The file names are load-bearing: the factory does lazy imports by name, and the
 
 ### Wiring
 
-Four places need to know about your policy. All by name.
+Two places need to know about your policy. All by name.
 
-1. **`policies/__init__.py`** — re-export `MyPolicyConfig` and add it to `__all__`. **Don't** re-export the modeling class; it loads lazily through the factory (so `import lerobot` stays fast).
-2. **`factory.py:get_policy_class`** — add a branch returning `MyPolicy` from a lazy import.
-3. **`factory.py:make_policy_config`** and **`factory.py:make_pre_post_processors`** — same idea, two more branches.
-4. **`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.
+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.
 
 Mirror an existing policy that's structurally similar to yours; the diff is small.
 
@@ -332,6 +330,10 @@ This way:
 
 Add a matching extra to [`pyproject.toml`](https://github.com/huggingface/lerobot/blob/main/pyproject.toml) `[project.optional-dependencies]` and include it in the `all` extra so `pip install 'lerobot[all]'` keeps installing everything.
 
+### Avoid copying a modeling file — subclass it
+
+If your policy needs to modify a backbone that already exists in `transformers` (custom conditioning, extra inputs, a swapped sub-module), **do not vendor a copy of its `modeling_*.py`**. Instead, subclass the smallest upstream unit and override only what changes. [`pi_gemma.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/pi_gemma.py) is the canonical reference: it injects AdaRMS conditioning into PaliGemma/Gemma in ~370 lines by subclassing `GemmaModel`/`PaliGemmaModel` and overriding the decoder-layer forward, instead of forking the ~2,000-line modeling file. Model surgery on a _loaded_ native model is also fine (layer truncation, tokenizer expansion, hidden-state capture — see `evo1/internvl3_embedder.py`, `eo1/modeling_eo1.py`, `groot/groot_n1_7.py` for working examples). Reviewers will ask for this pattern when a PR arrives with a copied modeling file; the only accepted exception is a model that does not exist in `transformers` at all.
+
 ### Benchmarks and a published checkpoint
 
 A new policy is much easier to review — and far more useful — when it ships with a working checkpoint and at least one number you can reproduce.
@@ -367,7 +369,7 @@ If your policy is real-robot-only and no sim benchmark applies, swap the sim eva
 The general expectations are in [`CONTRIBUTING.md`](https://github.com/huggingface/lerobot/blob/main/CONTRIBUTING.md) and the [PR template](https://github.com/huggingface/lerobot/blob/main/.github/PULL_REQUEST_TEMPLATE.md). On top of those, reviewers will look for:
 
 - [ ] `MyPolicy` and `MyPolicyConfig` cover the surface above; `__init_subclass__` accepts the class.
-- [ ] `factory.py` and `policies/__init__.py` are wired (lazy imports for modeling).
+- [ ] `policies/__init__.py` re-exports the config (this registers the policy; the factory resolves modeling/processor by naming convention).
 - [ ] `make_my_policy_pre_post_processors` follows the naming convention.
 - [ ] Optional deps live behind a `[project.optional-dependencies]` extra and the `TYPE_CHECKING + require_package` guard.
 - [ ] `tests/policies/` updated; backward-compat artifact committed & policy-specific tests.

examples/annotations/run_hf_job.py

@@ -46,8 +46,11 @@ CMD = (
     "apt-get update -qq && apt-get install -y -qq git ffmpeg && "
     "pip install --no-deps "
     "'lerobot @ git+https://github.com/huggingface/lerobot.git@main' && "
+    # Pins mirror pyproject.toml — unpinned installs pull av 18 / datasets 5 /
+    # draccus 0.11, which break lerobot at import time.
     "pip install --upgrade-strategy only-if-needed "
-    "datasets pyarrow av jsonlines draccus gymnasium torchcodec mergedeep pyyaml-include toml typing-inspect "
+    "'datasets>=4.7.0,<5.0.0' 'pyarrow>=21.0.0,<30.0.0' 'av>=15.0.0,<16.0.0' 'draccus==0.10.0' "
+    "'pandas>=2.0.0,<3.0.0' jsonlines gymnasium torchcodec mergedeep pyyaml-include toml typing-inspect "
     "openai && "
     "export VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=0 && "
     "export VLLM_VIDEO_BACKEND=pyav && "

src/lerobot/annotations/steerable_pipeline/config.py

@@ -65,6 +65,14 @@ class PlanConfig:
     # invented from the task text (+1 VLM call/episode).
     subtask_describe_first: bool = True
 
+    # Seeded relabeling: after segmentation, re-label each span with a focused
+    # pass that sees the previous / current / next segment contact sheets and
+    # minimally corrects the seed label (macrodata's best end-to-end labeling
+    # step). Costs +1 VLM call per subtask; off by default.
+    subtask_seeded_relabel: bool = False
+    # Frames sampled uniformly per segment sheet in the relabel pass.
+    subtask_relabel_frames: int = 5
+
     # Emit ``style="plan"`` rows at each boundary; False = subtasks + memory only.
     emit_plan: bool = True
 
@@ -160,6 +168,11 @@ class VlmConfig:
     # Forwarded as extra_body.chat_template_kwargs (e.g. {"enable_thinking": false}).
     chat_template_kwargs: dict[str, Any] | None = None
 
+    # OpenAI-style thinking budget hint ("low"/"medium"/"high"); forwarded to
+    # the server when set. Used to cap a thinking model's reasoning so it
+    # leaves tokens for the actual JSON answer on OpenAI-compatible endpoints.
+    reasoning_effort: str | None = None
+
 
 @dataclass
 class ExecutorConfig:

src/lerobot/annotations/steerable_pipeline/frames.py

@@ -413,7 +413,16 @@ def _draw_timestamp_badge(image: PIL.Image.Image, timestamp: float) -> PIL.Image
 
     result = image.copy()
     draw = ImageDraw.Draw(result)
-    font = ImageFont.load_default()
+    # Scale the timestamp to the tile so it stays legible after the model
+    # downsamples the full sheet into 768px tiles — a tiny bitmap font blurs
+    # at contact-sheet resolution and the VLM can no longer read the exact
+    # source time, which is what the boundary score depends on. ``size=`` is
+    # supported by Pillow's bitmap default since 10.1; fall back otherwise.
+    badge_px = max(14, round(image.height * 0.12))
+    try:
+        font = ImageFont.load_default(size=badge_px)
+    except TypeError:
+        font = ImageFont.load_default()
     label = f"{timestamp:06.2f}s"
     left, top, right, bottom = draw.textbbox((0, 0), label, font=font)
     text_w, text_h = right - left, bottom - top

src/lerobot/annotations/steerable_pipeline/modules/plan_subtasks_memory.py

@@ -116,6 +116,8 @@ class PlanSubtasksMemoryModule:
             rows.extend(self._task_aug_rows([effective_task, *variants], t0))
 
         subtask_spans = self._generate_subtasks(record, task=effective_task)
+        if self.config.subtask_seeded_relabel and subtask_spans:
+            subtask_spans = self._seeded_relabel(record, subtask_spans, effective_task)
 
         # subtask rows
         for span in subtask_spans:
@@ -509,6 +511,51 @@ class PlanSubtasksMemoryModule:
 
         return cleaned
 
+    def _seeded_relabel(
+        self, record: EpisodeRecord, spans: list[dict[str, Any]], task: str
+    ) -> list[dict[str, Any]]:
+        """Re-label each span using prev/current/next segment contact sheets.
+
+        Boundaries are kept fixed; only ``text`` is refined. The original
+        ("seed") label is passed as a strong prior so the model verifies and
+        minimally corrects it rather than re-describing from scratch — the
+        macrodata seeded-relabeling step. One VLM call per span.
+        """
+        n = len(spans)
+        out: list[dict[str, Any]] = []
+        for i, span in enumerate(spans):
+            content: list[dict[str, Any]] = []
+            if i > 0:
+                content += self._segment_sheet(record, spans[i - 1])
+            content += self._segment_sheet(record, span)
+            if i < n - 1:
+                content += self._segment_sheet(record, spans[i + 1])
+            prompt = load_prompt("plan_subtask_relabel").format(
+                episode_task=task,
+                seed_label=span["text"],
+                segment_index=i + 1,
+                segment_count=n,
+                start=float(span["start"]),
+                end=float(span["end"]),
+            )
+            content.append({"type": "text", "text": prompt})
+            label = self._vlm_field([{"role": "user", "content": content}], "label")
+            text = label.strip() if isinstance(label, str) and label.strip() else span["text"]
+            out.append({**span, "text": text})
+        return out
+
+    def _segment_sheet(self, record: EpisodeRecord, span: dict[str, Any]) -> list[dict[str, Any]]:
+        """Contact-sheet block(s) for one span: up to N frames sampled uniformly."""
+        s, e = float(span["start"]), float(span["end"])
+        n = max(1, int(self.config.subtask_relabel_frames))
+        if e <= s or n == 1:
+            timestamps = [s]
+        else:
+            step = (e - s) / (n - 1)
+            timestamps = [s + i * step for i in range(n)]
+        frames = self.frame_provider.frames_at(record, timestamps)
+        return self._contact_sheet_blocks(frames, timestamps[: len(frames)])
+
     def _generate_subtasks_windowed(
         self, record: EpisodeRecord, task: str, window_s: float
     ) -> list[dict[str, Any]]:

src/lerobot/annotations/steerable_pipeline/prompts/__init__.py

@@ -22,12 +22,23 @@ plain editors and roundtrip cleanly through ``ruff format``.
 
 from __future__ import annotations
 
+import os
 from pathlib import Path
 
 _DIR = Path(__file__).parent
 
 
 def load(name: str) -> str:
-    """Read prompt template ``name.txt`` from the ``prompts/`` directory."""
+    """Read prompt template ``name.txt`` from the ``prompts/`` directory.
+
+    A ``LEROBOT_PROMPT_OVERRIDE_<name>`` environment variable, when set to a
+    non-empty value, takes precedence over the packaged file. This lets prompt
+    search (e.g. GEPA) inject candidate templates into a remote job without
+    rebuilding the package; the override must keep the same ``{placeholder}``
+    fields the call site formats in.
+    """
+    override = os.environ.get(f"LEROBOT_PROMPT_OVERRIDE_{name}")
+    if override and override.strip():
+        return override
     path = _DIR / f"{name}.txt"
     return path.read_text(encoding="utf-8")

src/lerobot/annotations/steerable_pipeline/prompts/plan_subtask_relabel.txt

@@ -0,0 +1,35 @@
+Annotate one fixed segment from a longer robot demonstration.
+
+Return only JSON:
+  {{"label": "<short descriptive subtask label>"}}
+
+You are shown up to three timestamped contact sheets, in order:
+- The FIRST sheet is the PREVIOUS segment (context only); it may be absent.
+- The SECOND sheet is the CURRENT target segment.
+- The THIRD sheet is the NEXT segment (context only); it may be absent.
+Each tile has its timestamp (seconds, absolute video time) burned into its
+top-left corner.
+
+Episode instruction: "{episode_task}"
+Target segment: {segment_index} of {segment_count}
+Target time: {start:.2f}s to {end:.2f}s
+Original predicted label for this exact segment: "{seed_label}"
+
+Rules:
+- Label ONLY the current target segment (the second sheet). Use the
+  previous/next sheets only to disambiguate what changed.
+- Treat the original predicted label as a STRONG PRIOR, not ground truth:
+  verify it against the current segment and correct it minimally.
+- If it already names the right action and main object, keep it; only fix
+  grammar or add a clearly visible essential detail.
+- If it is vague but directionally correct, make it more specific.
+- If it describes the previous/next segment, the wrong action, wrong
+  object, wrong destination, or a wrong state change, replace it.
+- Do not describe the previous or next segment, and do not split, merge,
+  or move the fixed segment.
+- Do not introduce an action that is not clearly visible in the current
+  target segment.
+- Use one concise imperative phrase. Name the manipulated object and the
+  action / state change. Include source, destination, side, direction,
+  final placement, or opened/closed state when visible and central.
+- Do not mention timestamps, frame numbers, uncertainty, or intent.

src/lerobot/annotations/steerable_pipeline/prompts/plan_subtasks.txt

@@ -1,112 +1,68 @@
-You are labeling a teleoperated robot demonstration.
+You are annotating a teleoperated robot demonstration shown as
+timestamped contact sheets (each tile has its time in seconds burned
+into the top-left corner). The operator's goal was: "{episode_task}"
 
-The user originally asked: "{episode_task}"
+{observation_block}Reconstruct the sequence of COMPLETED manipulation events the robot
+performs, in chronological order. Output one segment per event with a
+[start, end] time in seconds and a short action label.
 
-You are shown the entire demonstration as a single video. Watch the
-whole clip, then segment it into a list of consecutive atomic subtasks
-the robot performs.
+GROUNDING — read first, it overrides everything below:
+- Label ONLY events you can SEE in the frames. The instruction is the
+  goal; the VIDEO is the ground truth for what actually happened.
+- Do NOT invent, anticipate, or pad steps that are not shown.
 
-{observation_block}GROUNDING — read this first, it overrides everything below:
-- Label ONLY what the robot actually does in the video. Every subtask
-  you emit must correspond to motion you can SEE in specific frames.
-- Do NOT invent, anticipate, or pad. If the robot only does one thing
-  (e.g. it just navigates to a location and the clip ends), emit
-  EXACTLY ONE subtask. Many demonstrations are a single atomic skill.
-- ``max_steps`` below is a hard CEILING, not a target. Emitting fewer
-  subtasks than the ceiling is not just allowed, it is expected for
-  short / atomic demonstrations. One correct subtask is far better
-  than several invented ones.
-- If the video does not clearly show the action implied by the task,
-  describe what you actually see — do NOT fabricate the task's steps
-  from the instruction text. The instruction tells you the goal; the
-  VIDEO is the ground truth for what happened.
+Granularity — segment by completed events, not by motion:
+- Start a NEW segment whenever the world state changes: an object is
+  grasped, lifted, transported, placed, or released; a held object
+  changes; a drawer/door/lid/container opens or closes; contents move
+  between containers (poured); a tool starts or stops acting on a
+  surface. Watch the gripper open/close transitions — they usually mark
+  boundaries.
+- Do NOT split approach, reach, grasp adjustment, small repositioning,
+  hesitation, or retreat into their own segments. Fold each into the
+  event it belongs to (the approach is part of the pick; the retreat is
+  part of the place).
+- Do NOT merge separate completed events. Each distinct pick, place,
+  open, close, pour, push, wipe, or insert is its own segment, even when
+  they repeat on different objects or locations.
+- Most segments last 2-10 seconds. Shorter segments are okay ONLY for
+  fast pick / place / open / close / release events. Never emit a
+  segment shorter than {min_subtask_seconds} seconds; merge a too-short
+  candidate into its neighbour instead.
+- Skip idle time, pure camera motion, and tiny hand jitter.
 
-Authoring rules — Hi Robot atom granularity, pi0.7-style short prompts:
+Labels — short imperative phrases:
+- One concise command naming the action and the manipulated object, e.g.
+  "pick up the red cup", "put the cup on the shelf", "open the top
+  drawer", "pour water into the glass", "insert the plug into the
+  socket".
+- Include source, destination, side, direction, or the final
+  open/closed state when it is visible and central to the event.
+- Prefer these verbs (extend only when none fits): pick up, put, place,
+  push, pull, turn, press, open, close, pour, insert, wipe, stack.
+  Disambiguate by what you SEE:
+    * STACK vs PUT: object placed ON TOP OF another object -> "stack".
+    * INSERT vs PUT: object pushed INTO a fitted slot/hole/socket -> "insert".
+    * PICK UP vs PUT (direction): gripper CLOSES and object moves WITH
+      the hand -> "pick up"; gripper OPENS and object stays -> "put".
+    * POUR vs PUT: source is tilted and contents flow -> "pour".
+- Use the exact object nouns implied by the task; stay consistent across
+  the episode (don't switch "cube" to "block").
+- Write imperative commands, never third person ("the robot ..."), and
+  drop articles/adverbs.
 
-- Each subtask = one COMPOSITE atomic skill the low-level policy can
-  execute end-to-end. A "skill" bundles its own approach motion with
-  its terminal action — do NOT split the approach off as its own
-  subtask. The whole-arm policy already learns to reach as part of
-  every manipulation primitive.
-- Write each subtask as an IMPERATIVE COMMAND, starting with one of
-  these verbs (extend only when none fits):
-    pick up <obj>           — approach + grasp + lift in one subtask
-    put <obj> on/in <loc>   — transport + release in one subtask
-    place <obj> on/in <loc> — synonym of "put"; pick one and stay consistent
-    push <obj>              — contact + linear shove
-    pull <obj>              — contact + linear retract
-    turn <knob/dial/handle> — rotary actuation
-    press <button>          — single-press contact
-    open <drawer/door/lid>  — full open motion
-    close <drawer/door/lid> — full close motion
-    pour <src> into <dst>   — tilt + flow
-    insert <obj> into <slot>— alignment + push-fit
-    go to <loc>             — ONLY when no grasp / actuation follows
-                             (e.g. a pure relocation between phases).
-                             If the next subtask grasps something at
-                             that location, drop "go to ..." and just
-                             write "pick up ..." instead.
-- Forbidden ultra-fine splits — the VLM is NOT allowed to emit these
-  as standalone subtasks; fold them into the parent composite:
-    "move to X"   → fold into "pick up X" (or whatever follows)
-    "reach for X" → fold into "pick up X"
-    "grasp X"     → fold into "pick up X"
-    "lift X"      → fold into "pick up X" (or "put X on Y" if it's
-                    the transport phase of a place)
-    "release X"   → fold into "put X on Y" (or "place X in Y")
-- Keep it SHORT — a verb phrase, not a sentence. Drop articles
-  ("the", "a") and adverbs ("carefully", "slowly"). Add a "how"
-  detail (which hand, which grasp point) ONLY when it is needed to
-  disambiguate. Every subtask must begin with one of the verbs
-  above (no leading nouns, no "then", no "first").
-- NEVER use third person. Never write "the robot", "the arm", "the
-  gripper moves", "it picks up" — the robot is implied. Command it,
-  do not describe it.
-- Use the exact object nouns from the task above. If the task says
-  "cube", every subtask says "cube" — never switch to "block". If it
-  says "box", never switch to "bin"/"container". Keep vocabulary
-  consistent across the whole episode.
-- Good: "pick up blue cube", "put blue cube in box", "open drawer",
-  "turn red knob", "press start button", "go to sink".
-- Bad: "move to blue cube" (approach as its own subtask — forbidden,
-  must be folded into "pick up blue cube"); "the robot arm moves
-  towards the blue cube" (third person, too long); "carefully pick
-  up the cube" (adverb, article); "release the yellow block"
-  ("block" when the task said "cube", and "release" must be folded
-  into a "put"/"place" subtask).
-- Subtasks are non-overlapping and cover the full episode in order.
-  Choose the cut points yourself based on what you see in the video
-  (gripper open/close events, contact, regrasps, transitions).
-- Each subtask spans at least {min_subtask_seconds} seconds. If a
-  candidate span would be shorter, merge it into its neighbour
-  rather than emitting it.
-- Do not exceed {max_steps} subtasks total. Fewer, larger composites
-  are preferred over many micro-steps.
-- Every subtask's [start_time, end_time] must lie within
-  [0.0, {episode_duration}] seconds.
-
-SPECIAL CASES — verb disambiguation (each rule is narrowly visual and
-fires ONLY on the spatial situation it names; it must not change how you
-label any other situation):
-- STACK vs PUT: if an object is placed ON TOP OF another specific object
-  (not on a flat table / shelf / counter), use "stack ... on ...", not
-  "put". "stack blue book on green book", NOT "put blue book on table".
-- INSERT vs PUT: if an object goes INTO a fitted slot / hole / socket /
-  receptacle (push-fit), use "insert ... into ...", not "put".
-- RETRIEVE/PICK-UP vs PUT (direction): watch the gripper. If it CLOSES
-  on the object and the object moves WITH the hand, it is "pick up" /
-  "retrieve" (object leaves its location). If the gripper OPENS and the
-  object stays where the hand left it, it is "put" / "place" (object
-  arrives at a location). Decide by which way the object moves, not by
-  where the hand ends up.
-- POUR vs PUT: only use "pour" when the source is tilted and contents
-  flow out; moving a full container without tilting is "put"/"place".
+Timing:
+- Use the burned-in timestamps to set start and end. Boundaries should
+  land on or near a printed time, and every [start, end] must lie within
+  [0.0, {episode_duration}] seconds, be non-overlapping, and cover the
+  episode in order.
+- Emit at most {max_steps} segments.
 
 Output strictly valid JSON of shape:
 
   {{
     "subtasks": [
-      {{"text": "<short imperative verb phrase>", "start": <float>, "end": <float>}},
+      {{"text": "<short imperative action label>", "start": <float>, "end": <float>}},
       ...
     ]
   }}

src/lerobot/annotations/steerable_pipeline/vlm_client.py

@@ -285,6 +285,8 @@ def _make_openai_client(config: VlmConfig) -> VlmClient:
             "max_tokens": max_tok,
             "temperature": temp,
         }
+        if config.reasoning_effort:
+            kwargs["reasoning_effort"] = config.reasoning_effort
         extra_body: dict[str, Any] = {}
         if send_mm_kwargs and mm_kwargs:
             extra_body["mm_processor_kwargs"] = {**mm_kwargs, "do_sample_frames": True}
@@ -296,7 +298,13 @@ def _make_openai_client(config: VlmConfig) -> VlmClient:
             chosen = clients[rr_counter["i"] % len(clients)]
             rr_counter["i"] += 1
         response = chosen.chat.completions.create(**kwargs)
-        return response.choices[0].message.content or ""
+        # Some OpenAI-compatible servers can return a choice with no message
+        # (safety filter, or a "thinking" model that spends the whole budget
+        # before emitting content). Treat that as an empty reply so the
+        # JSON-retry path handles it instead of crashing the run.
+        choice = response.choices[0] if response.choices else None
+        message = choice.message if choice is not None else None
+        return (message.content if message is not None else None) or ""
 
     def _gen(batch: Sequence[Sequence[dict[str, Any]]], max_tok: int, temp: float) -> list[str]:
         if len(batch) <= 1 or config.client_concurrency <= 1:

src/lerobot/configs/policies.py

@@ -205,24 +205,30 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno
                     f"{CONFIG_NAME} not found on the HuggingFace Hub in {model_id}"
                 ) from e
 
-        # HACK: Parse the original config to get the config subclass, so that we can
-        # apply cli overrides.
-        # This is very ugly, ideally we'd like to be able to do that natively with draccus
-        # something like --policy.path (in addition to --policy.type)
-        with draccus.config_type("json"):
-            orig_config = draccus.parse(cls, config_file, args=[])
-
         if config_file is None:
             raise FileNotFoundError(f"{CONFIG_NAME} not found in {model_id}")
 
         with open(config_file) as f:
             config = json.load(f)
 
-        config.pop("type")
+        # Resolve the concrete config subclass from the serialized "type" tag, then parse
+        # the config (with CLI overrides) directly for that class. The "type" key is
+        # stripped because draccus only consumes it when parsing the registry base class.
+        policy_type = config.pop("type", None)
+        if policy_type is None:
+            raise ValueError(f"Missing 'type' field in {CONFIG_NAME} of {model_id}")
+        try:
+            config_cls = cls.get_choice_class(policy_type)
+        except Exception as e:
+            raise ValueError(
+                f"Policy type '{policy_type}' (from {CONFIG_NAME} of {model_id}) is not registered. "
+                f"Available policy types: {cls.get_known_choices()}"
+            ) from e
+
         with tempfile.NamedTemporaryFile("w+", delete=False, suffix=".json") as f:
             json.dump(config, f)
             config_file = f.name
 
         cli_overrides = policy_kwargs.pop("cli_overrides", [])
         with draccus.config_type("json"):
-            return draccus.parse(orig_config.__class__, config_file, args=cli_overrides)
+            return draccus.parse(config_cls, config_file, args=cli_overrides)

src/lerobot/policies/__init__.py

@@ -32,6 +32,7 @@ from .pretrained import PreTrainedPolicy as PreTrainedPolicy
 from .smolvla.configuration_smolvla import SmolVLAConfig as SmolVLAConfig
 from .tdmpc.configuration_tdmpc import TDMPCConfig as TDMPCConfig
 from .utils import make_robot_action, prepare_observation_for_inference
+from .vla_jepa.configuration_vla_jepa import VLAJEPAConfig as VLAJEPAConfig
 from .vqbet.configuration_vqbet import VQBeTConfig as VQBeTConfig
 from .wall_x.configuration_wall_x import WallXConfig as WallXConfig
 from .xvla.configuration_xvla import XVLAConfig as XVLAConfig
@@ -57,6 +58,7 @@ __all__ = [
     "PI05Config",
     "SmolVLAConfig",
     "TDMPCConfig",
+    "VLAJEPAConfig",
     "VQBeTConfig",
     "WallXConfig",
     "XVLAConfig",

src/lerobot/policies/act/processor_act.py

@@ -18,17 +18,10 @@ from typing import Any
 import torch
 
 from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
-    RenameObservationsProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
+    make_default_pre_post_processors,
 )
-from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 from .configuration_act import ACTConfig
 
@@ -54,34 +47,4 @@ def make_act_pre_post_processors(
         tuple[PolicyProcessorPipeline[dict[str, Any], dict[str, Any]], PolicyProcessorPipeline[PolicyAction, PolicyAction]]: A tuple containing the
         pre-processor pipeline and the post-processor pipeline.
     """
-
-    input_steps = [
-        RenameObservationsProcessorStep(rename_map={}),
-        AddBatchDimensionProcessorStep(),
-        DeviceProcessorStep(device=config.device),
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-            device=config.device,
-        ),
-    ]
-    output_steps = [
-        UnnormalizerProcessorStep(
-            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
-        ),
-        DeviceProcessorStep(device="cpu"),
-    ]
-
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_default_pre_post_processors(config, dataset_stats, normalizer_device=config.device)

src/lerobot/policies/diffusion/processor_diffusion.py

@@ -19,17 +19,10 @@ from typing import Any
 import torch
 
 from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
-    RenameObservationsProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
+    make_default_pre_post_processors,
 )
-from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 from .configuration_diffusion import DiffusionConfig
 
@@ -63,32 +56,4 @@ def make_diffusion_pre_post_processors(
     Returns:
         A tuple containing the configured pre-processor and post-processor pipelines.
     """
-
-    input_steps = [
-        RenameObservationsProcessorStep(rename_map={}),
-        AddBatchDimensionProcessorStep(),
-        DeviceProcessorStep(device=config.device),
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
-    ]
-    output_steps = [
-        UnnormalizerProcessorStep(
-            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
-        ),
-        DeviceProcessorStep(device="cpu"),
-    ]
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_default_pre_post_processors(config, dataset_stats)

src/lerobot/policies/eo1/processor_eo1.py

@@ -23,24 +23,16 @@ import torch
 
 from lerobot.configs.types import FeatureType, PipelineFeatureType, PolicyFeature
 from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
     ComplementaryDataProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
     ProcessorStep,
     ProcessorStepRegistry,
-    RenameObservationsProcessorStep,
-    UnnormalizerProcessorStep,
+    make_default_policy_processor_steps,
+    make_policy_processor_pipelines,
 )
-from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
 from lerobot.types import TransitionKey
-from lerobot.utils.constants import (
-    OBS_STATE,
-    POLICY_POSTPROCESSOR_DEFAULT_NAME,
-    POLICY_PREPROCESSOR_DEFAULT_NAME,
-)
+from lerobot.utils.constants import OBS_STATE
 from lerobot.utils.import_utils import _transformers_available, require_package
 
 from .configuration_eo1 import EO1Config
@@ -242,14 +234,12 @@ def make_eo1_pre_post_processors(
 ]:
     """Build pre/post processor pipelines for EO1."""
 
+    steps = make_default_policy_processor_steps(config, dataset_stats)
+
     input_steps: list[ProcessorStep] = [
-        RenameObservationsProcessorStep(rename_map={}),
-        AddBatchDimensionProcessorStep(),
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
+        steps.rename_observations,
+        steps.add_batch_dim,
+        steps.normalize,
         EO1ConversationTemplateStep(input_features=config.input_features, chunk_size=config.chunk_size),
         EO1QwenProcessorStep(
             processor_name=config.vlm_base,
@@ -257,27 +247,12 @@ def make_eo1_pre_post_processors(
             image_max_pixels=config.image_max_pixels,
             use_fast_processor=config.use_fast_processor,
         ),
-        DeviceProcessorStep(device=config.device),
+        steps.to_device,
     ]
 
     output_steps: list[ProcessorStep] = [
-        UnnormalizerProcessorStep(
-            features=config.output_features,
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
-        DeviceProcessorStep(device="cpu"),
+        steps.unnormalize,
+        steps.to_cpu,
     ]
 
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_policy_processor_pipelines(input_steps=input_steps, output_steps=output_steps)

src/lerobot/policies/evo1/internvl3_embedder.py

@@ -27,9 +27,11 @@ from lerobot.utils.import_utils import _transformers_available, require_package
 
 if TYPE_CHECKING or _transformers_available:
     from transformers import AutoModel, AutoTokenizer
+    from transformers.utils import is_flash_attn_2_available
 else:
     AutoModel = None
     AutoTokenizer = None
+    is_flash_attn_2_available = None
 
 IMAGENET_MEAN = (0.485, 0.456, 0.406)
 IMAGENET_STD = (0.229, 0.224, 0.225)
@@ -135,9 +137,13 @@ class InternVL3Embedder(nn.Module):
                 raise ValueError(f"Unsupported EVO1 vlm_dtype '{model_dtype}'") from exc
         self.model_dtype = model_dtype
 
-        attn_implementation = "flash_attention_2" if (use_flash_attn and _flash_attn_available()) else "eager"
+        attn_implementation = (
+            "flash_attention_2" if (use_flash_attn and is_flash_attn_2_available()) else "eager"
+        )
         if use_flash_attn and attn_implementation == "eager":
-            logger.warning("flash_attn is not installed. Falling back to eager attention.")
+            logger.warning(
+                "Flash Attention 2 is unavailable on this runtime. Falling back to eager attention."
+            )
 
         self.model = AutoModel.from_pretrained(
             model_name,
@@ -359,11 +365,3 @@ class InternVL3Embedder(nn.Module):
     @property
     def device(self) -> torch.device:
         return next(self.model.parameters()).device
-
-
-def _flash_attn_available() -> bool:
-    try:
-        import flash_attn  # noqa: F401
-    except ModuleNotFoundError:
-        return False
-    return True

src/lerobot/policies/factory.py

@@ -17,6 +17,7 @@
 from __future__ import annotations
 
 import importlib
+import inspect
 import logging
 from typing import TYPE_CHECKING, Any, TypedDict, Unpack
 
@@ -44,26 +45,10 @@ from lerobot.utils.constants import (
 )
 from lerobot.utils.feature_utils import dataset_to_policy_features
 
-from .act.configuration_act import ACTConfig
-from .diffusion.configuration_diffusion import DiffusionConfig
-from .eo1.configuration_eo1 import EO1Config
 from .evo1.configuration_evo1 import Evo1Config
-from .fastwam.configuration_fastwam import FastWAMConfig
-from .gaussian_actor.configuration_gaussian_actor import GaussianActorConfig
 from .groot.configuration_groot import GrootConfig
-from .lingbot_va.configuration_lingbot_va import LingBotVAConfig
-from .molmoact2.configuration_molmoact2 import MolmoAct2Config
-from .multi_task_dit.configuration_multi_task_dit import MultiTaskDiTConfig
-from .pi0.configuration_pi0 import PI0Config
-from .pi05.configuration_pi05 import PI05Config
 from .pretrained import PreTrainedPolicy
-from .smolvla.configuration_smolvla import SmolVLAConfig
-from .tdmpc.configuration_tdmpc import TDMPCConfig
 from .utils import validate_visual_features_consistency
-from .vla_jepa.configuration_vla_jepa import VLAJEPAConfig
-from .vqbet.configuration_vqbet import VQBeTConfig
-from .wall_x.configuration_wall_x import WallXConfig
-from .xvla.configuration_xvla import XVLAConfig
 
 
 def _reconnect_relative_absolute_steps(
@@ -88,100 +73,23 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
     """
     Retrieves a policy class by its registered name.
 
-    This function uses dynamic imports to avoid loading all policy classes into memory
-    at once, improving startup time and reducing dependencies.
+    Resolution is convention-based: the draccus-registered config class of ``name`` is
+    looked up, its ``configuration_*`` module path is rewritten to ``modeling_*``, and
+    the ``<X>Policy`` class is imported from there. The modeling module is only imported
+    at call time, keeping heavy optional dependencies lazy. This works for both built-in
+    policies and third-party lerobot plugins (anything registered via
+    ``@PreTrainedConfig.register_subclass``).
 
     Args:
-        name: The name of the policy. Supported names are "tdmpc", "diffusion", "act",
-            "multi_task_dit", "vqbet", "pi0", "pi05", "gaussian_actor", "smolvla", "wall_x",
-            "molmoact2", "eo1", "evo1".
+        name: The registered name of the policy (e.g. "act", "diffusion", "pi0").
     Returns:
         The policy class corresponding to the given name.
 
     Raises:
-        NotImplementedError: If the policy name is not recognized.
+        ValueError: If the policy name is not registered.
+        ImportError: If the policy's optional dependencies are not installed.
     """
-    if name == "tdmpc":
-        from .tdmpc.modeling_tdmpc import TDMPCPolicy
-
-        return TDMPCPolicy
-    elif name == "diffusion":
-        from .diffusion.modeling_diffusion import DiffusionPolicy
-
-        return DiffusionPolicy
-    elif name == "act":
-        from .act.modeling_act import ACTPolicy
-
-        return ACTPolicy
-    elif name == "multi_task_dit":
-        from .multi_task_dit.modeling_multi_task_dit import MultiTaskDiTPolicy
-
-        return MultiTaskDiTPolicy
-    elif name == "vqbet":
-        from .vqbet.modeling_vqbet import VQBeTPolicy
-
-        return VQBeTPolicy
-    elif name == "pi0":
-        from .pi0.modeling_pi0 import PI0Policy
-
-        return PI0Policy
-    elif name == "pi0_fast":
-        from .pi0_fast.modeling_pi0_fast import PI0FastPolicy
-
-        return PI0FastPolicy
-    elif name == "pi05":
-        from .pi05.modeling_pi05 import PI05Policy
-
-        return PI05Policy
-    elif name == "gaussian_actor":
-        from .gaussian_actor.modeling_gaussian_actor import GaussianActorPolicy
-
-        return GaussianActorPolicy
-    elif name == "smolvla":
-        from .smolvla.modeling_smolvla import SmolVLAPolicy
-
-        return SmolVLAPolicy
-    elif name == "groot":
-        from .groot.modeling_groot import GrootPolicy
-
-        return GrootPolicy
-    elif name == "xvla":
-        from .xvla.modeling_xvla import XVLAPolicy
-
-        return XVLAPolicy
-    elif name == "wall_x":
-        from .wall_x.modeling_wall_x import WallXPolicy
-
-        return WallXPolicy
-    elif name == "eo1":
-        from .eo1.modeling_eo1 import EO1Policy
-
-        return EO1Policy
-    elif name == "molmoact2":
-        from .molmoact2.modeling_molmoact2 import MolmoAct2Policy
-
-        return MolmoAct2Policy
-    elif name == "vla_jepa":
-        from .vla_jepa.modeling_vla_jepa import VLAJEPAPolicy
-
-        return VLAJEPAPolicy
-    elif name == "lingbot_va":
-        from .lingbot_va.modeling_lingbot_va import LingBotVAPolicy
-
-        return LingBotVAPolicy
-    elif name == "fastwam":
-        from .fastwam.modeling_fastwam import FastWAMPolicy
-
-        return FastWAMPolicy
-    elif name == "evo1":
-        from .evo1.modeling_evo1 import Evo1Policy
-
-        return Evo1Policy
-    else:
-        try:
-            return _get_policy_cls_from_policy_name(name=name)
-        except Exception as e:
-            raise ValueError(f"Policy type '{name}' is not available.") from e
+    return _get_policy_cls_from_policy_name(name=name)
 
 
 def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
@@ -192,9 +100,8 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
     mapping a string identifier to the corresponding config class.
 
     Args:
-        policy_type: The type of the policy. Supported types include "tdmpc",
-                     "multi_task_dit", "diffusion", "act", "vqbet", "pi0", "pi05", "gaussian_actor",
-                     "smolvla", "wall_x", "molmoact2", "eo1", "evo1".
+        policy_type: The registered type of the policy (any name registered via
+                     ``@PreTrainedConfig.register_subclass``, e.g. "act", "diffusion", "pi0").
         **kwargs: Keyword arguments to be passed to the configuration class constructor.
 
     Returns:
@@ -203,48 +110,11 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
     Raises:
         ValueError: If the `policy_type` is not recognized.
     """
-    if policy_type == "tdmpc":
-        return TDMPCConfig(**kwargs)
-    elif policy_type == "diffusion":
-        return DiffusionConfig(**kwargs)
-    elif policy_type == "act":
-        return ACTConfig(**kwargs)
-    elif policy_type == "multi_task_dit":
-        return MultiTaskDiTConfig(**kwargs)
-    elif policy_type == "vqbet":
-        return VQBeTConfig(**kwargs)
-    elif policy_type == "pi0":
-        return PI0Config(**kwargs)
-    elif policy_type == "pi05":
-        return PI05Config(**kwargs)
-    elif policy_type == "gaussian_actor":
-        return GaussianActorConfig(**kwargs)
-    elif policy_type == "smolvla":
-        return SmolVLAConfig(**kwargs)
-    elif policy_type == "groot":
-        return GrootConfig(**kwargs)
-    elif policy_type == "xvla":
-        return XVLAConfig(**kwargs)
-    elif policy_type == "wall_x":
-        return WallXConfig(**kwargs)
-    elif policy_type == "eo1":
-        return EO1Config(**kwargs)
-    elif policy_type == "molmoact2":
-        return MolmoAct2Config(**kwargs)
-    elif policy_type == "vla_jepa":
-        return VLAJEPAConfig(**kwargs)
-    elif policy_type == "lingbot_va":
-        return LingBotVAConfig(**kwargs)
-    elif policy_type == "fastwam":
-        return FastWAMConfig(**kwargs)
-    elif policy_type == "evo1":
-        return Evo1Config(**kwargs)
-    else:
-        try:
-            config_cls = PreTrainedConfig.get_choice_class(policy_type)
-            return config_cls(**kwargs)
-        except Exception as e:
-            raise ValueError(f"Policy type '{policy_type}' is not available.") from e
+    try:
+        config_cls = PreTrainedConfig.get_choice_class(policy_type)
+    except Exception as e:
+        raise ValueError(f"Policy type '{policy_type}' is not available.") from e
+    return config_cls(**kwargs)
 
 
 class ProcessorConfigKwargs(TypedDict, total=False):
@@ -298,8 +168,7 @@ def make_pre_post_processors(
         A tuple containing the input (pre-processor) and output (post-processor) pipelines.
 
     Raises:
-        NotImplementedError: If a processor factory is not implemented for the given
-            policy configuration type.
+        ValueError: If no processor factory exists for the given policy configuration type.
     """
     if pretrained_path:
         if isinstance(policy_cfg, GrootConfig):
@@ -351,166 +220,13 @@ def make_pre_post_processors(
             )
         return preprocessor, postprocessor
 
-    # Create a new processor based on policy type
-    if isinstance(policy_cfg, TDMPCConfig):
-        from .tdmpc.processor_tdmpc import make_tdmpc_pre_post_processors
-
-        processors = make_tdmpc_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, DiffusionConfig):
-        from .diffusion.processor_diffusion import make_diffusion_pre_post_processors
-
-        processors = make_diffusion_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, ACTConfig):
-        from .act.processor_act import make_act_pre_post_processors
-
-        processors = make_act_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, MultiTaskDiTConfig):
-        from .multi_task_dit.processor_multi_task_dit import (
-            make_multi_task_dit_pre_post_processors,
-        )
-
-        processors = make_multi_task_dit_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, VQBeTConfig):
-        from .vqbet.processor_vqbet import make_vqbet_pre_post_processors
-
-        processors = make_vqbet_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, PI0Config):
-        from .pi0.processor_pi0 import make_pi0_pre_post_processors
-
-        processors = make_pi0_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, PI05Config):
-        from .pi05.processor_pi05 import make_pi05_pre_post_processors
-
-        processors = make_pi05_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, GaussianActorConfig):
-        from .gaussian_actor.processor_gaussian_actor import make_gaussian_actor_pre_post_processors
-
-        processors = make_gaussian_actor_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, SmolVLAConfig):
-        from .smolvla.processor_smolvla import make_smolvla_pre_post_processors
-
-        processors = make_smolvla_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, GrootConfig):
-        from .groot.processor_groot import make_groot_pre_post_processors
-
-        processors = make_groot_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-            dataset_meta=kwargs.get("dataset_meta"),
-        )
-
-    elif isinstance(policy_cfg, XVLAConfig):
-        from .xvla.processor_xvla import (
-            make_xvla_pre_post_processors,
-        )
-
-        processors = make_xvla_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, WallXConfig):
-        from .wall_x.processor_wall_x import make_wall_x_pre_post_processors
-
-        processors = make_wall_x_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, EO1Config):
-        from .eo1.processor_eo1 import make_eo1_pre_post_processors
-
-        processors = make_eo1_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-    elif isinstance(policy_cfg, Evo1Config):
-        from .evo1.processor_evo1 import make_evo1_pre_post_processors
-
-        processors = make_evo1_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, MolmoAct2Config):
-        from .molmoact2.processor_molmoact2 import make_molmoact2_pre_post_processors
-
-        processors = make_molmoact2_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-            dataset_meta=kwargs.get("dataset_meta"),
-        )
-
-    elif isinstance(policy_cfg, VLAJEPAConfig):
-        from .vla_jepa.processor_vla_jepa import make_vla_jepa_pre_post_processors
-
-        processors = make_vla_jepa_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, LingBotVAConfig):
-        from .lingbot_va.processor_lingbot_va import make_lingbot_va_pre_post_processors
-
-        processors = make_lingbot_va_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, FastWAMConfig):
-        from .fastwam.processor_fastwam import make_fastwam_pre_post_processors
-
-        processors = make_fastwam_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    else:
-        try:
-            processors = _make_processors_from_policy_config(
-                config=policy_cfg,
-                dataset_stats=kwargs.get("dataset_stats"),
-            )
-        except Exception as e:
-            raise ValueError(f"Processor for policy type '{policy_cfg.type}' is not implemented.") from e
-
-    return processors
+    # Create new processors from the policy config, resolving the per-policy factory
+    # function by naming convention (lazy import keeps optional dependencies optional).
+    return _make_processors_from_policy_config(
+        config=policy_cfg,
+        dataset_stats=kwargs.get("dataset_stats"),
+        dataset_meta=kwargs.get("dataset_meta"),
+    )
 
 
 def make_policy(
@@ -654,10 +370,12 @@ def make_policy(
     return policy
 
 
-def _get_policy_cls_from_policy_name(name: str) -> type[PreTrainedConfig]:
+def _get_policy_cls_from_policy_name(name: str) -> type[PreTrainedPolicy]:
     """Get policy class from its registered name using dynamic imports.
 
-    This is used as a helper function to import policies from 3rd party lerobot plugins.
+    Works for built-in policies and 3rd party lerobot plugins alike: the config class
+    registered under ``name`` is resolved via the draccus ChoiceRegistry, and the policy
+    class is imported from the sibling ``modeling_*`` module by naming convention.
 
     Args:
         name: The name of the policy.
@@ -683,22 +401,39 @@ def _get_policy_cls_from_policy_name(name: str) -> type[PreTrainedConfig]:
         "configuration_", "modeling_"
     )  # e.g., configuration_diffusion -> modeling_diffusion
 
-    module = importlib.import_module(module_path)
-    policy_cls = getattr(module, cls_name)
+    try:
+        module = importlib.import_module(module_path)
+    except ModuleNotFoundError as e:
+        if e.name == module_path:
+            # The modeling_* module itself does not exist for this policy type. A missing
+            # optional dependency inside an existing module propagates unchanged instead,
+            # so its actionable install hint stays visible.
+            raise ValueError(f"Policy class for '{name}' is not implemented.") from e
+        raise
+    policy_cls = getattr(module, cls_name, None)
+    if policy_cls is None:
+        raise ValueError(
+            f"Policy class '{cls_name}' not found in '{module_path}'. "
+            f"Policies must expose '<Name>Policy' in the sibling 'modeling_*' module by naming convention."
+        )
     return policy_cls
 
 
 def _make_processors_from_policy_config(
     config: PreTrainedConfig,
     dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+    dataset_meta: Any | None = None,
 ) -> tuple[Any, Any]:
     """Create pre- and post-processors from a policy configuration using dynamic imports.
 
-    This is used as a helper function to import processor factories from 3rd party lerobot plugins.
+    Resolves ``make_{type}_pre_post_processors`` from the policy's ``processor_*`` module
+    by naming convention. Works for built-in policies and 3rd party lerobot plugins.
 
     Args:
         config: The policy configuration object.
         dataset_stats: Dataset statistics for normalization.
+        dataset_meta: Dataset metadata, forwarded only to factories that declare a
+            ``dataset_meta`` parameter (e.g. groot, molmoact2).
     Returns:
         A tuple containing the input (pre-processor) and output (post-processor) pipelines.
     """
@@ -711,6 +446,19 @@ def _make_processors_from_policy_config(
     logging.debug(
         f"Instantiating pre/post processors using function '{function_name}' from module '{module_path}'"
     )
-    module = importlib.import_module(module_path)
-    function = getattr(module, function_name)
-    return function(config, dataset_stats=dataset_stats)
+    try:
+        module = importlib.import_module(module_path)
+    except ModuleNotFoundError as e:
+        if e.name == module_path:
+            # The processor_* module itself does not exist for this policy type. A missing
+            # optional dependency inside an existing module propagates unchanged instead,
+            # so its actionable install hint stays visible.
+            raise ValueError(f"Processor for policy type '{policy_type}' is not implemented.") from e
+        raise
+    function = getattr(module, function_name, None)
+    if function is None:
+        raise ValueError(f"Processor for policy type '{policy_type}' is not implemented.")
+    call_kwargs: dict[str, Any] = {"dataset_stats": dataset_stats}
+    if "dataset_meta" in inspect.signature(function).parameters:
+        call_kwargs["dataset_meta"] = dataset_meta
+    return function(config, **call_kwargs)

src/lerobot/policies/fastwam/processor_fastwam.py

@@ -22,20 +22,11 @@ import torch
 from lerobot.configs import PipelineFeatureType, PolicyFeature
 from lerobot.processor import (
     ActionProcessorStep,
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
     ProcessorStepRegistry,
-    RenameObservationsProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
-)
-from lerobot.utils.constants import (
-    POLICY_POSTPROCESSOR_DEFAULT_NAME,
-    POLICY_PREPROCESSOR_DEFAULT_NAME,
+    make_default_policy_processor_steps,
+    make_policy_processor_pipelines,
 )
 
 from .configuration_fastwam import FastWAMConfig
@@ -105,38 +96,20 @@ def make_fastwam_pre_post_processors(
     # anyway) and unsafe across fine-tuning: its `resize_size` would be inherited from the base
     # checkpoint's camera geometry, not this dataset's, making the concatenation N_cameras x too wide.
 
+    steps = make_default_policy_processor_steps(config, normalization_stats, normalizer_device=config.device)
+
     input_steps = [
-        RenameObservationsProcessorStep(rename_map={}),
-        AddBatchDimensionProcessorStep(),
-        DeviceProcessorStep(device=config.device),
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=normalization_stats,
-            device=config.device,
-        ),
+        steps.rename_observations,
+        steps.add_batch_dim,
+        steps.to_device,
+        steps.normalize,
     ]
     output_steps = [
-        UnnormalizerProcessorStep(
-            features=config.output_features,
-            norm_map=config.normalization_mapping,
-            stats=normalization_stats,
-        ),
+        steps.unnormalize,
     ]
     if config.toggle_action_dimensions:
         output_steps.append(
             FastWAMActionToggleProcessorStep(toggle_dimensions=config.toggle_action_dimensions)
         )
-    output_steps.append(DeviceProcessorStep(device="cpu"))
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    output_steps.append(steps.to_cpu)
+    return make_policy_processor_pipelines(input_steps=input_steps, output_steps=output_steps)

src/lerobot/policies/gaussian_actor/processor_gaussian_actor.py

@@ -20,17 +20,10 @@ from typing import Any
 import torch
 
 from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
-    RenameObservationsProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
+    make_default_pre_post_processors,
 )
-from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 from .configuration_gaussian_actor import GaussianActorConfig
 
@@ -62,33 +55,4 @@ def make_gaussian_actor_pre_post_processors(
     Returns:
         A tuple containing the configured pre-processor and post-processor pipelines.
     """
-
-    # Add remaining processors
-    input_steps = [
-        RenameObservationsProcessorStep(rename_map={}),
-        AddBatchDimensionProcessorStep(),
-        DeviceProcessorStep(device=config.device),
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
-    ]
-    output_steps = [
-        UnnormalizerProcessorStep(
-            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
-        ),
-        DeviceProcessorStep(device="cpu"),
-    ]
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_default_pre_post_processors(config, dataset_stats)

src/lerobot/policies/lingbot_va/processor_lingbot_va.py

@@ -25,19 +25,12 @@ import torch
 
 from lerobot.configs.types import FeatureType, NormalizationMode
 from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
     ProcessorStep,
-    RenameObservationsProcessorStep,
     UnnormalizerProcessorStep,
-)
-from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
-from lerobot.utils.constants import (
-    POLICY_POSTPROCESSOR_DEFAULT_NAME,
-    POLICY_PREPROCESSOR_DEFAULT_NAME,
+    make_default_policy_processor_steps,
+    make_policy_processor_pipelines,
 )
 
 from .configuration_lingbot_va import LingBotVAConfig
@@ -52,15 +45,13 @@ def make_lingbot_va_pre_post_processors(
 ]:
     """Build the pre/post processor pipelines for LingBot-VA."""
 
+    steps = make_default_policy_processor_steps(config, dataset_stats)
+
     input_steps: list[ProcessorStep] = [
-        RenameObservationsProcessorStep(rename_map={}),
-        AddBatchDimensionProcessorStep(),
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
-        DeviceProcessorStep(device=config.device),
+        steps.rename_observations,
+        steps.add_batch_dim,
+        steps.normalize,
+        steps.to_device,
     ]
 
     # Unnormalize actions from [-1, 1] to physical units (QUANTILES) using q01/q99 restored from the checkpoint.
@@ -70,18 +61,7 @@ def make_lingbot_va_pre_post_processors(
             norm_map={FeatureType.ACTION: NormalizationMode.QUANTILES},
             stats=dataset_stats,
         ),
-        DeviceProcessorStep(device="cpu"),
+        steps.to_cpu,
     ]
 
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_policy_processor_pipelines(input_steps=input_steps, output_steps=output_steps)

src/lerobot/policies/multi_task_dit/processor_multi_task_dit.py

@@ -19,18 +19,12 @@ from typing import Any
 import torch
 
 from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
-    RenameObservationsProcessorStep,
     TokenizerProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
+    make_default_policy_processor_steps,
+    make_policy_processor_pipelines,
 )
-from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 from .configuration_multi_task_dit import MultiTaskDiTConfig
 
@@ -66,9 +60,11 @@ def make_multi_task_dit_pre_post_processors(
         A tuple containing the configured pre-processor and post-processor pipelines.
     """
 
+    steps = make_default_policy_processor_steps(config, dataset_stats, normalizer_device=config.device)
+
     input_steps = [
-        RenameObservationsProcessorStep(rename_map={}),
-        AddBatchDimensionProcessorStep(),
+        steps.rename_observations,
+        steps.add_batch_dim,
         TokenizerProcessorStep(
             tokenizer_name=config.text_encoder_name,
             padding=config.tokenizer_padding,
@@ -76,32 +72,12 @@ def make_multi_task_dit_pre_post_processors(
             max_length=config.tokenizer_max_length,
             truncation=config.tokenizer_truncation,
         ),
-        DeviceProcessorStep(device=config.device),
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-            device=config.device,
-        ),
+        steps.to_device,
+        steps.normalize,
     ]
     output_steps = [
-        UnnormalizerProcessorStep(
-            features=config.output_features,
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
-        DeviceProcessorStep(device="cpu"),
+        steps.unnormalize,
+        steps.to_cpu,
     ]
 
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_policy_processor_pipelines(input_steps=input_steps, output_steps=output_steps)

src/lerobot/policies/pi0/processor_pi0.py

@@ -21,22 +21,16 @@ import torch
 from lerobot.configs import PipelineFeatureType, PolicyFeature
 from lerobot.processor import (
     AbsoluteActionsProcessorStep,
-    AddBatchDimensionProcessorStep,
     ComplementaryDataProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
     ProcessorStep,
     ProcessorStepRegistry,
     RelativeActionsProcessorStep,
-    RenameObservationsProcessorStep,
     TokenizerProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
+    make_default_policy_processor_steps,
+    make_policy_processor_pipelines,
 )
-from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 from .configuration_pi0 import PI0Config
 
@@ -136,10 +130,12 @@ def make_pi0_pre_post_processors(
         action_names=getattr(config, "action_feature_names", None),
     )
 
+    steps = make_default_policy_processor_steps(config, dataset_stats)
+
     # OpenPI order: raw → relative → normalize → model → unnormalize → absolute
     input_steps: list[ProcessorStep] = [
-        RenameObservationsProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
-        AddBatchDimensionProcessorStep(),
+        steps.rename_observations,  # To mimic the same processor as pretrained one
+        steps.add_batch_dim,
         Pi0NewLineProcessor(),  # Add newlines before tokenization for PaliGemma
         TokenizerProcessorStep(
             tokenizer_name="google/paligemma-3b-pt-224",
@@ -147,32 +143,15 @@ def make_pi0_pre_post_processors(
             padding_side="right",
             padding="max_length",
         ),
-        DeviceProcessorStep(device=config.device),
+        steps.to_device,
         relative_step,
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
+        steps.normalize,
     ]
 
     output_steps: list[ProcessorStep] = [
-        UnnormalizerProcessorStep(
-            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
-        ),
+        steps.unnormalize,
         AbsoluteActionsProcessorStep(enabled=config.use_relative_actions, relative_step=relative_step),
-        DeviceProcessorStep(device="cpu"),
+        steps.to_cpu,
     ]
 
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_policy_processor_pipelines(input_steps=input_steps, output_steps=output_steps)

src/lerobot/policies/pi05/processor_pi05.py

@@ -24,26 +24,17 @@ import torch
 from lerobot.configs import PipelineFeatureType, PolicyFeature
 from lerobot.processor import (
     AbsoluteActionsProcessorStep,
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
     ProcessorStep,
     ProcessorStepRegistry,
     RelativeActionsProcessorStep,
-    RenameObservationsProcessorStep,
     TokenizerProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
+    make_default_policy_processor_steps,
+    make_policy_processor_pipelines,
 )
 from lerobot.types import EnvTransition, TransitionKey
-from lerobot.utils.constants import (
-    OBS_STATE,
-    POLICY_POSTPROCESSOR_DEFAULT_NAME,
-    POLICY_PREPROCESSOR_DEFAULT_NAME,
-)
+from lerobot.utils.constants import OBS_STATE
 
 from .configuration_pi05 import PI05Config
 
@@ -135,18 +126,16 @@ def make_pi05_pre_post_processors(
         action_names=getattr(config, "action_feature_names", None),
     )
 
+    steps = make_default_policy_processor_steps(config, dataset_stats)
+
     # OpenPI order: raw → relative → normalize → model → unnormalize → absolute
     input_steps: list[ProcessorStep] = [
-        RenameObservationsProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
-        AddBatchDimensionProcessorStep(),
+        steps.rename_observations,  # To mimic the same processor as pretrained one
+        steps.add_batch_dim,
         relative_step,
         # NOTE: NormalizerProcessorStep MUST come before Pi05PrepareStateTokenizerProcessorStep
         # because the tokenizer step expects normalized state in [-1, 1] range for discretization
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
+        steps.normalize,
         Pi05PrepareStateTokenizerProcessorStep(max_state_dim=config.max_state_dim),
         TokenizerProcessorStep(
             tokenizer_name="google/paligemma-3b-pt-224",
@@ -154,26 +143,13 @@ def make_pi05_pre_post_processors(
             padding_side="right",
             padding="max_length",
         ),
-        DeviceProcessorStep(device=config.device),
+        steps.to_device,
     ]
 
     output_steps: list[ProcessorStep] = [
-        UnnormalizerProcessorStep(
-            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
-        ),
+        steps.unnormalize,
         AbsoluteActionsProcessorStep(enabled=config.use_relative_actions, relative_step=relative_step),
-        DeviceProcessorStep(device="cpu"),
+        steps.to_cpu,
     ]
 
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_policy_processor_pipelines(input_steps=input_steps, output_steps=output_steps)

src/lerobot/policies/pi0_fast/processor_pi0_fast.py

@@ -25,26 +25,17 @@ from lerobot.configs import PipelineFeatureType, PolicyFeature
 from lerobot.processor import (
     AbsoluteActionsProcessorStep,
     ActionTokenizerProcessorStep,
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
     ProcessorStep,
     ProcessorStepRegistry,
     RelativeActionsProcessorStep,
-    RenameObservationsProcessorStep,
     TokenizerProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
+    make_default_policy_processor_steps,
+    make_policy_processor_pipelines,
 )
 from lerobot.types import EnvTransition, TransitionKey
-from lerobot.utils.constants import (
-    OBS_STATE,
-    POLICY_POSTPROCESSOR_DEFAULT_NAME,
-    POLICY_PREPROCESSOR_DEFAULT_NAME,
-)
+from lerobot.utils.constants import OBS_STATE
 
 from .configuration_pi0_fast import PI0FastConfig
 
@@ -135,6 +126,8 @@ def make_pi0_fast_pre_post_processors(
         action_names=getattr(config, "action_feature_names", None),
     )
 
+    steps = make_default_policy_processor_steps(config, dataset_stats)
+
     # Pi0Fast order: relative → normalize → tokenize → model → unnormalize → absolute
     # This matches pi0/pi0.5: RelativeActionsProcessorStep runs first on raw absolute actions,
     # caching the raw state. NormalizerProcessorStep then normalizes the raw relative actions,
@@ -144,14 +137,10 @@ def make_pi0_fast_pre_post_processors(
     # before Pi0FastPrepareStateAndLanguageTokenizerProcessorStep, so the state tokenizer
     # continues to receive normalized state in [-1, 1] as expected.
     input_steps: list[ProcessorStep] = [
-        RenameObservationsProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
-        AddBatchDimensionProcessorStep(),
+        steps.rename_observations,  # To mimic the same processor as pretrained one
+        steps.add_batch_dim,
         relative_step,
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
+        steps.normalize,
         Pi0FastPrepareStateAndLanguageTokenizerProcessorStep(max_state_dim=config.max_state_dim),
         TokenizerProcessorStep(
             tokenizer_name=config.text_tokenizer_name,
@@ -165,26 +154,13 @@ def make_pi0_fast_pre_post_processors(
             fast_skip_tokens=config.fast_skip_tokens,
             paligemma_tokenizer_name=config.text_tokenizer_name,
         ),
-        DeviceProcessorStep(device=config.device),
+        steps.to_device,
     ]
 
     output_steps: list[ProcessorStep] = [
-        UnnormalizerProcessorStep(
-            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
-        ),
+        steps.unnormalize,
         AbsoluteActionsProcessorStep(enabled=config.use_relative_actions, relative_step=relative_step),
-        DeviceProcessorStep(device="cpu"),
+        steps.to_cpu,
     ]
 
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_policy_processor_pipelines(input_steps=input_steps, output_steps=output_steps)

src/lerobot/policies/pretrained.py

@@ -23,8 +23,6 @@ from pathlib import Path
 from tempfile import TemporaryDirectory
 from typing import TYPE_CHECKING, TypedDict, TypeVar, Unpack
 
-import packaging
-import safetensors
 from huggingface_hub import HfApi, ModelCard, ModelCardData, hf_hub_download, save_torch_state_dict
 from huggingface_hub.constants import SAFETENSORS_SINGLE_FILE
 from huggingface_hub.errors import HfHubHTTPError
@@ -34,6 +32,7 @@ from torch import Tensor, nn
 from lerobot.__version__ import __version__
 from lerobot.configs import PreTrainedConfig
 from lerobot.configs.train import TrainPipelineConfig
+from lerobot.utils.device_utils import resolve_safetensors_device
 from lerobot.utils.hub import HubMixin
 
 from .utils import log_model_loading_keys
@@ -221,26 +220,10 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
 
     @classmethod
     def _load_as_safetensor(cls, model: T, model_file: str, map_location: str, strict: bool) -> T:
-        # Create base kwargs
-        kwargs = {"strict": strict}
-
-        # Add device parameter for newer versions that support it
-        if packaging.version.parse(safetensors.__version__) >= packaging.version.parse("0.4.3"):
-            kwargs["device"] = map_location
-
-        # Load the model with appropriate kwargs
-        missing_keys, unexpected_keys = load_model_as_safetensor(model, model_file, **kwargs)
+        missing_keys, unexpected_keys = load_model_as_safetensor(
+            model, model_file, strict=strict, device=resolve_safetensors_device(map_location)
+        )
         log_model_loading_keys(missing_keys, unexpected_keys)
-
-        # For older versions, manually move to device if needed
-        if "device" not in kwargs and map_location != "cpu":
-            logging.warning(
-                "Loading model weights on other devices than 'cpu' is not supported natively in your version of safetensors."
-                " This means that the model is loaded on 'cpu' first and then copied to the device."
-                " This leads to a slower loading time."
-                " Please update safetensors to version 0.4.3 or above for improved performance."
-            )
-            model.to(map_location)
         return model
 
     @abc.abstractmethod

src/lerobot/policies/smolvla/processor_smolvla.py

@@ -19,19 +19,13 @@ from typing import Any
 import torch
 
 from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
     NewLineTaskProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
-    RenameObservationsProcessorStep,
     TokenizerProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
+    make_default_policy_processor_steps,
+    make_policy_processor_pipelines,
 )
-from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 from .configuration_smolvla import SmolVLAConfig
 
@@ -66,9 +60,11 @@ def make_smolvla_pre_post_processors(
         A tuple containing the configured pre-processor and post-processor pipelines.
     """
 
+    steps = make_default_policy_processor_steps(config, dataset_stats)
+
     input_steps = [
-        RenameObservationsProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
-        AddBatchDimensionProcessorStep(),
+        steps.rename_observations,  # To mimic the same processor as pretrained one
+        steps.add_batch_dim,
         NewLineTaskProcessorStep(),
         TokenizerProcessorStep(
             tokenizer_name=config.vlm_model_name,
@@ -76,28 +72,11 @@ def make_smolvla_pre_post_processors(
             padding_side="right",
             max_length=config.tokenizer_max_length,
         ),
-        DeviceProcessorStep(device=config.device),
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
+        steps.to_device,
+        steps.normalize,
     ]
     output_steps = [
-        UnnormalizerProcessorStep(
-            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
-        ),
-        DeviceProcessorStep(device="cpu"),
+        steps.unnormalize,
+        steps.to_cpu,
     ]
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_policy_processor_pipelines(input_steps=input_steps, output_steps=output_steps)

src/lerobot/policies/tdmpc/processor_tdmpc.py

@@ -19,17 +19,10 @@ from typing import Any
 import torch
 
 from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
-    RenameObservationsProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
+    make_default_pre_post_processors,
 )
-from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 from .configuration_tdmpc import TDMPCConfig
 
@@ -61,32 +54,4 @@ def make_tdmpc_pre_post_processors(
     Returns:
         A tuple containing the configured pre-processor and post-processor pipelines.
     """
-
-    input_steps = [
-        RenameObservationsProcessorStep(rename_map={}),
-        AddBatchDimensionProcessorStep(),
-        DeviceProcessorStep(device=config.device),
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
-    ]
-    output_steps = [
-        UnnormalizerProcessorStep(
-            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
-        ),
-        DeviceProcessorStep(device="cpu"),
-    ]
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_default_pre_post_processors(config, dataset_stats)

src/lerobot/policies/vla_jepa/processor_vla_jepa.py

@@ -20,20 +20,16 @@ import torch
 
 from lerobot.policies.vla_jepa.configuration_vla_jepa import VLAJEPAConfig
 from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
     EnvTransition,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
     ProcessorStep,
     ProcessorStepRegistry,
-    RenameObservationsProcessorStep,
     TransitionKey,
     UnnormalizerProcessorStep,
+    make_default_policy_processor_steps,
+    make_policy_processor_pipelines,
 )
-from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
-from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 
 @ProcessorStepRegistry.register(name="vla_jepa_clip_actions")
@@ -112,15 +108,12 @@ def make_vla_jepa_pre_post_processors(
     PolicyProcessorPipeline[PolicyAction, PolicyAction],
 ]:
     features = {**config.input_features, **config.output_features}
+    steps = make_default_policy_processor_steps(config, dataset_stats)
     input_steps = [
-        RenameObservationsProcessorStep(rename_map={}),
-        AddBatchDimensionProcessorStep(),
-        DeviceProcessorStep(device=config.device),
-        NormalizerProcessorStep(
-            features=features,
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
+        steps.rename_observations,
+        steps.add_batch_dim,
+        steps.to_device,
+        steps.normalize,
     ]
     output_steps: list[ProcessorStep] = []
     if config.clip_normalized_actions:
@@ -129,6 +122,8 @@ def make_vla_jepa_pre_post_processors(
         output_steps.append(
             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.
     output_steps.append(
         UnnormalizerProcessorStep(
             features=features,
@@ -140,16 +135,5 @@ def make_vla_jepa_pre_post_processors(
         output_steps.append(
             BinarizeGripperProcessorStep(gripper_dim=config.gripper_dim, threshold=config.gripper_threshold)
         )
-    output_steps.append(DeviceProcessorStep(device="cpu"))
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    output_steps.append(steps.to_cpu)
+    return make_policy_processor_pipelines(input_steps=input_steps, output_steps=output_steps)

src/lerobot/policies/vqbet/processor_vqbet.py

@@ -20,17 +20,10 @@ from typing import Any
 import torch
 
 from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
-    RenameObservationsProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
+    make_default_pre_post_processors,
 )
-from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 from .configuration_vqbet import VQBeTConfig
 
@@ -62,32 +55,4 @@ def make_vqbet_pre_post_processors(
     Returns:
         A tuple containing the configured pre-processor and post-processor pipelines.
     """
-
-    input_steps = [
-        RenameObservationsProcessorStep(rename_map={}),  # Let the possibility to the user to rename the keys
-        AddBatchDimensionProcessorStep(),
-        DeviceProcessorStep(device=config.device),
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
-    ]
-    output_steps = [
-        UnnormalizerProcessorStep(
-            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
-        ),
-        DeviceProcessorStep(device="cpu"),
-    ]
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_default_pre_post_processors(config, dataset_stats)

src/lerobot/policies/wall_x/processor_wall_x.py

@@ -20,19 +20,13 @@ import torch
 
 from lerobot.configs import PipelineFeatureType, PolicyFeature
 from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
     ComplementaryDataProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
     ProcessorStepRegistry,
-    RenameObservationsProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
+    make_default_policy_processor_steps,
+    make_policy_processor_pipelines,
 )
-from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 from .configuration_wall_x import WallXConfig
 
@@ -65,37 +59,22 @@ def make_wall_x_pre_post_processors(
         A tuple containing the configured pre-processor and post-processor pipelines
     """
 
+    steps = make_default_policy_processor_steps(config, dataset_stats)
+
     input_steps = [
-        RenameObservationsProcessorStep(rename_map={}),
-        AddBatchDimensionProcessorStep(),
+        steps.rename_observations,
+        steps.add_batch_dim,
         WallXTaskProcessor(),  # Process task description
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
-        DeviceProcessorStep(device=config.device),
+        steps.normalize,
+        steps.to_device,
     ]
 
     output_steps = [
-        UnnormalizerProcessorStep(
-            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
-        ),
-        DeviceProcessorStep(device="cpu"),
+        steps.unnormalize,
+        steps.to_cpu,
     ]
 
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_policy_processor_pipelines(input_steps=input_steps, output_steps=output_steps)
 
 
 @ProcessorStepRegistry.register(name="wall_x_task_processor")

src/lerobot/policies/xvla/processor_xvla.py

@@ -22,19 +22,14 @@ import torch
 
 from lerobot.configs import PipelineFeatureType, PolicyFeature
 from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     ObservationProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
     ProcessorStep,
     ProcessorStepRegistry,
-    RenameObservationsProcessorStep,
     TokenizerProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
+    make_default_policy_processor_steps,
+    make_policy_processor_pipelines,
 )
 from lerobot.types import EnvTransition, TransitionKey
 from lerobot.utils.constants import (
@@ -42,8 +37,6 @@ from lerobot.utils.constants import (
     OBS_IMAGES,
     OBS_PREFIX,
     OBS_STATE,
-    POLICY_POSTPROCESSOR_DEFAULT_NAME,
-    POLICY_PREPROCESSOR_DEFAULT_NAME,
 )
 
 from .configuration_xvla import XVLAConfig
@@ -61,10 +54,11 @@ def make_xvla_pre_post_processors(
     Build the LeRobot processor pipelines for XVLA.
     """
 
-    features = {**config.input_features, **config.output_features}
+    steps = make_default_policy_processor_steps(config, dataset_stats)
+
     input_steps = [
-        RenameObservationsProcessorStep(rename_map={}),
-        AddBatchDimensionProcessorStep(),
+        steps.rename_observations,
+        steps.add_batch_dim,
         TokenizerProcessorStep(
             tokenizer_name=config.tokenizer_name,
             max_length=config.tokenizer_max_length,
@@ -74,32 +68,15 @@ def make_xvla_pre_post_processors(
         XVLAImageToFloatProcessorStep(),
         XVLAImageNetNormalizeProcessorStep(),
         XVLAAddDomainIdProcessorStep(),
-        DeviceProcessorStep(device=config.device),
-        NormalizerProcessorStep(
-            features=features, norm_map=config.normalization_mapping, stats=dataset_stats
-        ),
+        steps.to_device,
+        steps.normalize,
     ]
     output_steps = [
-        UnnormalizerProcessorStep(
-            features=config.output_features,
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
-        DeviceProcessorStep(device="cpu"),
+        steps.unnormalize,
+        steps.to_cpu,
     ]
 
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_policy_processor_pipelines(input_steps=input_steps, output_steps=output_steps)
 
 
 # Custom XVLA processor steps

src/lerobot/processor/__init__.py

@@ -42,10 +42,14 @@ from .delta_action_processor import MapDeltaActionToRobotActionStep, MapTensorTo
 from .device_processor import DeviceProcessorStep
 from .env_processor import IsaaclabArenaProcessorStep, LiberoProcessorStep
 from .factory import (
+    DefaultPolicyProcessorSteps,
+    make_default_policy_processor_steps,
+    make_default_pre_post_processors,
     make_default_processors,
     make_default_robot_action_processor,
     make_default_robot_observation_processor,
     make_default_teleop_action_processor,
+    make_policy_processor_pipelines,
 )
 from .gym_action_processor import (
     Numpy2TorchActionProcessorStep,
@@ -129,10 +133,14 @@ __all__ = [
     "ImageCropResizeProcessorStep",
     "InfoProcessorStep",
     "InterventionActionProcessorStep",
+    "DefaultPolicyProcessorSteps",
+    "make_default_policy_processor_steps",
+    "make_default_pre_post_processors",
     "make_default_processors",
     "make_default_teleop_action_processor",
     "make_default_robot_action_processor",
     "make_default_robot_observation_processor",
+    "make_policy_processor_pipelines",
     "AbsoluteActionsProcessorStep",
     "RelativeActionsProcessorStep",
     "MapDeltaActionToRobotActionStep",

src/lerobot/processor/factory.py

@@ -14,15 +14,33 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from lerobot.types import RobotAction, RobotObservation
+from dataclasses import dataclass
+from typing import Any
 
+import torch
+
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.types import PolicyAction, RobotAction, RobotObservation
+from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
+
+from .batch_processor import AddBatchDimensionProcessorStep
 from .converters import (
     observation_to_transition,
+    policy_action_to_transition,
     robot_action_observation_to_transition,
     transition_to_observation,
+    transition_to_policy_action,
     transition_to_robot_action,
 )
-from .pipeline import IdentityProcessorStep, RobotProcessorPipeline
+from .device_processor import DeviceProcessorStep
+from .normalize_processor import NormalizerProcessorStep, UnnormalizerProcessorStep
+from .pipeline import (
+    IdentityProcessorStep,
+    PolicyProcessorPipeline,
+    ProcessorStep,
+    RobotProcessorPipeline,
+)
+from .rename_processor import RenameObservationsProcessorStep
 
 
 def make_default_teleop_action_processor() -> RobotProcessorPipeline[
@@ -61,3 +79,97 @@ def make_default_processors():
     robot_action_processor = make_default_robot_action_processor()
     robot_observation_processor = make_default_robot_observation_processor()
     return (teleop_action_processor, robot_action_processor, robot_observation_processor)
+
+
+@dataclass
+class DefaultPolicyProcessorSteps:
+    """The canonical processor steps shared by most policies' pre/post pipelines.
+
+    Policies compose these in their own order (step ORDER is a Hub-serialized contract
+    and intentionally stays explicit per policy) and interleave their custom steps.
+    """
+
+    rename_observations: RenameObservationsProcessorStep
+    add_batch_dim: AddBatchDimensionProcessorStep
+    to_device: DeviceProcessorStep
+    normalize: NormalizerProcessorStep
+    unnormalize: UnnormalizerProcessorStep
+    to_cpu: DeviceProcessorStep
+
+
+def make_default_policy_processor_steps(
+    config: PreTrainedConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+    *,
+    normalizer_device: torch.device | str | None = None,
+) -> DefaultPolicyProcessorSteps:
+    """Construct the canonical policy processor steps from a policy config.
+
+    Args:
+        config: A `PreTrainedConfig` providing `device`, `input_features`,
+            `output_features` and `normalization_mapping`.
+        dataset_stats: Dataset statistics used for (un)normalization.
+        normalizer_device: Device passed to `NormalizerProcessorStep` (some policies pin
+            their normalization stats to the policy device; most leave it unset).
+    """
+    return DefaultPolicyProcessorSteps(
+        rename_observations=RenameObservationsProcessorStep(rename_map={}),
+        add_batch_dim=AddBatchDimensionProcessorStep(),
+        to_device=DeviceProcessorStep(device=config.device),
+        normalize=NormalizerProcessorStep(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+            device=normalizer_device,
+        ),
+        unnormalize=UnnormalizerProcessorStep(
+            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
+        ),
+        to_cpu=DeviceProcessorStep(device="cpu"),
+    )
+
+
+def make_policy_processor_pipelines(
+    input_steps: list[ProcessorStep],
+    output_steps: list[ProcessorStep],
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """Wrap pre/post step lists into the canonical policy pipeline pair.
+
+    Uses the standard pipeline names (which determine the serialized JSON filenames on
+    the Hub) and the standard policy-action converters on the postprocessor.
+    """
+    return (
+        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
+            steps=input_steps,
+            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+        ),
+        PolicyProcessorPipeline[PolicyAction, PolicyAction](
+            steps=output_steps,
+            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+            to_transition=policy_action_to_transition,
+            to_output=transition_to_policy_action,
+        ),
+    )
+
+
+def make_default_pre_post_processors(
+    config: PreTrainedConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+    *,
+    normalizer_device: torch.device | str | None = None,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """The pure-scaffold policy pipeline pair: Rename -> Batch -> Device -> Normalize,
+    and Unnormalize -> Device(cpu). Policies with custom steps or a different step order
+    compose `make_default_policy_processor_steps` themselves instead.
+    """
+    s = make_default_policy_processor_steps(config, dataset_stats, normalizer_device=normalizer_device)
+    return make_policy_processor_pipelines(
+        input_steps=[s.rename_observations, s.add_batch_dim, s.to_device, s.normalize],
+        output_steps=[s.unnormalize, s.to_cpu],
+    )

src/lerobot/processor/relative_action_processor.py

@@ -126,7 +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
 
@@ -146,6 +146,11 @@ 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,

src/lerobot/rewards/pretrained.py

@@ -21,8 +21,6 @@ from pathlib import Path
 from tempfile import TemporaryDirectory
 from typing import TYPE_CHECKING, Any, TypeVar
 
-import packaging
-import safetensors
 from huggingface_hub import HfApi, ModelCard, ModelCardData, hf_hub_download
 from huggingface_hub.constants import SAFETENSORS_SINGLE_FILE
 from huggingface_hub.errors import HfHubHTTPError
@@ -30,6 +28,7 @@ from safetensors.torch import load_model as load_model_as_safetensor, save_model
 from torch import Tensor, nn
 
 from lerobot.configs.rewards import RewardModelConfig
+from lerobot.utils.device_utils import resolve_safetensors_device
 from lerobot.utils.hub import HubMixin
 
 if TYPE_CHECKING:
@@ -129,29 +128,13 @@ class PreTrainedRewardModel(nn.Module, HubMixin, abc.ABC):
 
     @classmethod
     def _load_as_safetensor(cls, model: T, model_file: str, map_location: str, strict: bool) -> T:
-        # Create base kwargs
-        kwargs = {"strict": strict}
-
-        # Add device parameter for newer versions that support it
-        if packaging.version.parse(safetensors.__version__) >= packaging.version.parse("0.4.3"):
-            kwargs["device"] = map_location
-
-        # Load the model with appropriate kwargs
-        missing_keys, unexpected_keys = load_model_as_safetensor(model, model_file, **kwargs)
+        missing_keys, unexpected_keys = load_model_as_safetensor(
+            model, model_file, strict=strict, device=resolve_safetensors_device(map_location)
+        )
         if missing_keys:
             logging.warning(f"Missing key(s) when loading model: {missing_keys}")
         if unexpected_keys:
             logging.warning(f"Unexpected key(s) when loading model: {unexpected_keys}")
-
-        # For older versions, manually move to device if needed
-        if "device" not in kwargs and map_location != "cpu":
-            logging.warning(
-                "Loading model weights on other devices than 'cpu' is not supported natively in your version of safetensors."
-                " This means that the model is loaded on 'cpu' first and then copied to the device."
-                " This leads to a slower loading time."
-                " Please update safetensors to version 0.4.3 or above for improved performance."
-            )
-            model.to(map_location)
         return model
 
     def get_optim_params(self):

src/lerobot/rollout/context.py

@@ -43,7 +43,6 @@ 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
@@ -52,7 +51,6 @@ from .configs import BaseStrategyConfig, DAggerStrategyConfig, RolloutConfig
 from .inference import (
     InferenceEngine,
     RTCInferenceConfig,
-    SyncInferenceConfig,
     create_inference_engine,
 )
 from .robot_wrapper import ThreadSafeRobot
@@ -399,15 +397,6 @@ 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)...",

src/lerobot/rollout/inference/sync.py

@@ -24,26 +24,21 @@ import torch
 
 from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.policies.utils import make_robot_action, prepare_observation_for_inference
-from lerobot.processor import PolicyProcessorPipeline
+from lerobot.processor import PolicyProcessorPipeline, RelativeActionsProcessorStep
 
 from .base import InferenceEngine
 
 logger = logging.getLogger(__name__)
 
 
-# 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``).
+# 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.
 
 
 class SyncInferenceEngine(InferenceEngine):
@@ -73,6 +68,31 @@ class SyncInferenceEngine(InferenceEngine):
         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")
+
         logger.info(
             "SyncInferenceEngine initialized (device=%s, action_keys=%d)",
             self._device,
@@ -85,6 +105,11 @@ 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:
@@ -93,6 +118,27 @@ 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
+
+    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_action(self, obs_frame: dict | None) -> torch.Tensor | None:
         """Run the full inference pipeline on ``obs_frame`` and return an action tensor."""
@@ -107,12 +153,25 @@ 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
         with torch.inference_mode(), autocast_ctx:
             observation = prepare_observation_for_inference(
                 observation, self._device, self._task, self._robot_type
             )
             observation = self._preprocessor(observation)
             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._postprocessor(action)
         action_tensor = action.squeeze(0).cpu()
 

src/lerobot/scripts/lerobot_annotate.py

@@ -28,7 +28,12 @@ For distributed runs, see ``examples/annotations/run_hf_job.py``.
 """
 
 import logging
+from contextlib import suppress
 from pathlib import Path
+from typing import TYPE_CHECKING
+
+from huggingface_hub import HfApi, snapshot_download
+from huggingface_hub.errors import RevisionNotFoundError
 
 from lerobot.annotations.steerable_pipeline.config import AnnotationPipelineConfig
 from lerobot.annotations.steerable_pipeline.executor import Executor
@@ -42,6 +47,12 @@ from lerobot.annotations.steerable_pipeline.validator import StagingValidator
 from lerobot.annotations.steerable_pipeline.vlm_client import make_vlm_client
 from lerobot.annotations.steerable_pipeline.writer import LanguageColumnsWriter
 from lerobot.configs import parser
+from lerobot.utils.import_utils import _datasets_available, require_package
+
+if TYPE_CHECKING or _datasets_available:
+    from lerobot.datasets.dataset_metadata import CODEBASE_VERSION
+    from lerobot.datasets.io_utils import load_info
+    from lerobot.datasets.utils import create_lerobot_dataset_card
 
 logger = logging.getLogger(__name__)
 
@@ -50,8 +61,6 @@ def _resolve_root(cfg: AnnotationPipelineConfig) -> Path:
     if cfg.root is not None:
         return Path(cfg.root)
     if cfg.repo_id is not None:
-        from huggingface_hub import snapshot_download
-
         return Path(snapshot_download(repo_id=cfg.repo_id, repo_type="dataset"))
     raise ValueError("Either --root or --repo_id must be provided.")
 
@@ -125,7 +134,7 @@ def _push_to_hub(root: Path, cfg: AnnotationPipelineConfig) -> None:
 
     Pushes to ``cfg.new_repo_id`` when set, otherwise back to ``cfg.repo_id``.
     """
-    from huggingface_hub import HfApi  # noqa: PLC0415
+    require_package("datasets", "dataset")
 
     repo_id = cfg.new_repo_id or cfg.repo_id
     commit_message = cfg.push_commit_message or "Add steerable annotations (lerobot-annotate)"
@@ -143,33 +152,26 @@ def _push_to_hub(root: Path, cfg: AnnotationPipelineConfig) -> None:
         repo_id=repo_id,
         repo_type="dataset",
         commit_message=commit_message,
-        ignore_patterns=[".annotate_staging/**", "**/.DS_Store"],
+        # README.md is excluded because when pushing to ``new_repo_id`` the
+        # source card's links (e.g. the visualize badge) would keep pointing
+        # at the source dataset; a fresh card is generated below instead.
+        ignore_patterns=[".annotate_staging/**", "**/.DS_Store", "README.md"],
     )
     print(f"[lerobot-annotate] uploaded to https://huggingface.co/datasets/{repo_id}", flush=True)
 
+    dataset_info = load_info(root)
+    card = create_lerobot_dataset_card(dataset_info=dataset_info, license="apache-2.0", repo_id=repo_id)
+    card.push_to_hub(repo_id=repo_id, repo_type="dataset")
+
     # Tag the upload with the codebase version. ``LeRobotDatasetMetadata``
     # resolves the dataset revision via ``get_safe_version`` which scans
     # for tags like ``v3.0``; without a tag it raises
     # ``RevisionNotFoundError``. Read the version straight from the
     # dataset's own ``meta/info.json`` so we tag whatever the writer
     # actually wrote (no accidental drift if the codebase floor moves).
-    from lerobot.datasets.dataset_metadata import CODEBASE_VERSION  # noqa: PLC0415
-
-    info_path = root / "meta" / "info.json"
-    version_tag = CODEBASE_VERSION
-    if info_path.exists():
-        try:
-            from lerobot.utils.io_utils import load_json  # noqa: PLC0415
-
-            info = load_json(info_path)
-            ds_version = info.get("codebase_version")
-            if isinstance(ds_version, str) and ds_version.startswith("v"):
-                version_tag = ds_version
-        except Exception as exc:  # noqa: BLE001
-            print(
-                f"[lerobot-annotate] could not read codebase_version from info.json ({exc}); falling back to {version_tag}",
-                flush=True,
-            )
+    version_tag = (
+        dataset_info.codebase_version if dataset_info.codebase_version.startswith("v") else CODEBASE_VERSION
+    )
     revision = getattr(commit_info, "oid", None)
     tag_kwargs = {
         "repo_id": repo_id,
@@ -180,10 +182,6 @@ def _push_to_hub(root: Path, cfg: AnnotationPipelineConfig) -> None:
         tag_kwargs["revision"] = revision
 
     try:
-        from contextlib import suppress  # noqa: PLC0415
-
-        from huggingface_hub.errors import RevisionNotFoundError  # noqa: PLC0415
-
         with suppress(RevisionNotFoundError):
             api.delete_tag(repo_id, tag=version_tag, repo_type="dataset")
         api.create_tag(**tag_kwargs)

src/lerobot/scripts/lerobot_train.py

@@ -171,6 +171,9 @@ def update_policy(
     train_metrics.update_s = time.perf_counter() - start_time
     if torch.cuda.is_available():
         train_metrics.gpu_mem_gb = torch.cuda.max_memory_allocated() / (1024**3)
+    # Aggregate the policy's scalar outputs for logging and rank-reduction across the log window.
+    if output_dict:
+        train_metrics.update_metrics(output_dict)
     return train_metrics, output_dict
 
 
@@ -572,7 +575,7 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
         batch = preprocessor(batch)
         train_tracker.dataloading_s = time.perf_counter() - start_time
 
-        train_tracker, output_dict = update_policy(
+        train_tracker, _ = update_policy(
             train_tracker,
             policy,
             batch,
@@ -605,9 +608,10 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
                     train_tracker.samples_per_s = effective_batch_size / step_time
                 logging.info(train_tracker)
                 if wandb_logger:
+                    # Policy sub-losses (latent_loss, action_loss, ...) are aggregated into the
+                    # tracker by update_policy, so to_dict() already carries their windowed,
+                    # rank-reduced averages — no per-step output_dict passthrough needed.
                     wandb_log_dict = train_tracker.to_dict()
-                    if output_dict:
-                        wandb_log_dict.update(output_dict)
                     # Log sample weighting statistics if enabled
                     if sample_weighter is not None:
                         weighter_stats = sample_weighter.get_stats()

src/lerobot/utils/device_utils.py

@@ -59,6 +59,20 @@ def get_safe_torch_device(try_device: str, log: bool = False) -> torch.device:
     return device
 
 
+def resolve_safetensors_device(map_location: str | torch.device) -> str:
+    """Resolve a device string for a safetensors load, working around a device-mapping quirk.
+
+    safetensors' load maps the bare string "cuda" to cuda:0 regardless of the current device
+    (unlike torch's .to("cuda"), which honors torch.cuda.current_device()). Under multi-GPU
+    accelerate/FSDP every rank would then load its weights onto GPU 0, OOMing it before sharding.
+    Resolve "cuda" to the concrete current-device index so each rank loads onto its own GPU.
+    """
+    map_location = str(map_location)
+    if map_location == "cuda" and torch.cuda.is_available():
+        return f"cuda:{torch.cuda.current_device()}"
+    return map_location
+
+
 def get_safe_dtype(dtype: torch.dtype, device: str | torch.device):
     """
     mps is currently not compatible with float64

src/lerobot/utils/logging_utils.py

@@ -104,6 +104,7 @@ class MetricsTracker:
         "episodes",
         "epochs",
         "accelerator",
+        "_caller_metrics",
     ]
 
     def __init__(
@@ -129,6 +130,9 @@ class MetricsTracker:
         self.episodes = self.samples / self._avg_samples_per_ep
         self.epochs = self.samples / self._num_frames
         self.accelerator = accelerator
+        # Meter names the caller registered up front. update_metrics() leaves these untouched, so a
+        # policy that echoes e.g. "loss" in its output dict can't clobber the aggregated meter.
+        self._caller_metrics: set[str] = set(self.metrics)
 
     def __getattr__(self, name: str) -> int | dict[str, AverageMeter] | AverageMeter | Any:
         if name in self.__dict__:
@@ -156,6 +160,21 @@ class MetricsTracker:
         self.episodes = self.samples / self._avg_samples_per_ep
         self.epochs = self.samples / self._num_frames
 
+    def update_metrics(self, values: dict[str, Any]) -> None:
+        """Accumulate a dict of scalar metrics, auto-registering a meter for each new key.
+
+        Non-numeric values and bools are ignored.
+        Caller-registered metrics (those passed to the constructor) are never overridden.
+        """
+        for name, value in values.items():
+            if isinstance(value, bool) or not isinstance(value, (int, float)):
+                continue
+            if name in self._caller_metrics:
+                continue
+            if name not in self.metrics:
+                self.metrics[name] = AverageMeter(name, ":.3f", reduction="mean")
+            self.metrics[name].update(float(value))
+
     def reduce_across_ranks(self) -> None:
         """
         Synchronises the running averages of every metric whose ``reduction`` is not ``"none"``

tests/annotations/test_modules.py

@@ -85,7 +85,7 @@ def _spy_responder(captured: list[list[dict[str, Any]]], reply: Any):
 def test_module1_plan_memory_subtask_smoke(fixture_dataset_root: Path, tmp_path: Path) -> None:
     vlm = make_canned_responder(
         {
-            "atomic subtasks": {
+            "COMPLETED manipulation events": {
                 "subtasks": [
                     {"text": "grasp the handle of the sponge", "start": 0.0, "end": 0.4},
                     {"text": "wipe the counter from left to right", "start": 0.4, "end": 0.8},
@@ -126,7 +126,7 @@ def test_module1_emit_memory_false_skips_memory_keeps_subtasks_and_plan(
     leaving subtask + plan generation intact — symmetric to ``emit_plan``."""
     vlm = make_canned_responder(
         {
-            "atomic subtasks": {
+            "COMPLETED manipulation events": {
                 "subtasks": [
                     {"text": "grasp the handle of the sponge", "start": 0.0, "end": 0.4},
                     {"text": "wipe the counter from left to right", "start": 0.4, "end": 0.8},
@@ -318,7 +318,7 @@ def test_module1_attaches_contact_sheets_to_subtask_prompt(
                     return block.get("text", "")
         return ""
 
-    subtask_calls = [m for m in captured if "atomic subtasks" in _prompt_text(m)]
+    subtask_calls = [m for m in captured if "COMPLETED manipulation events" in _prompt_text(m)]
     assert len(subtask_calls) == 1, "expected exactly one subtask-prompt VLM call"
     content = subtask_calls[0][0]["content"]
     video_blocks = [b for b in content if isinstance(b, dict) and b.get("type") == "video"]

tests/policies/test_relative_actions.py

@@ -346,3 +346,10 @@ def test_state_not_modified_by_relative_processor(dataset, action_dim):
 
     result_state = result[TransitionKey.OBSERVATION][OBS_STATE]
     torch.testing.assert_close(result_state, original_state)
+
+
+def test_cached_anchor_not_in_config():
+    """The cached anchor is ephemeral runtime state and must not leak into the config."""
+    step = RelativeActionsProcessorStep(enabled=True)
+    step.set_cached_state(torch.tensor([[1.0, 2.0, 3.0, 4.0]]))
+    assert set(step.get_config()) == {"enabled", "exclude_joints", "action_names"}

tests/scripts/test_lerobot_annotate.py

@@ -18,6 +18,8 @@ import json
 from types import SimpleNamespace
 
 import pytest
+import requests
+from huggingface_hub.errors import RevisionNotFoundError
 
 # ``lerobot.scripts.lerobot_annotate`` (and the ``_push_to_hub`` path it
 # exercises) imports ``lerobot.datasets``, which only ships under the
@@ -26,11 +28,13 @@ pytest.importorskip("datasets", reason="datasets is required (install lerobot[da
 
 
 def test_push_to_hub_tags_uploaded_dataset_revision(tmp_path, monkeypatch):
-    from lerobot.scripts.lerobot_annotate import _push_to_hub
+    from lerobot.scripts import lerobot_annotate
 
     root = tmp_path / "dataset"
     (root / "meta").mkdir(parents=True)
-    (root / "meta" / "info.json").write_text(json.dumps({"codebase_version": "v3.0"}))
+    (root / "meta" / "info.json").write_text(
+        json.dumps({"codebase_version": "v3.0", "fps": 30, "features": {}})
+    )
 
     calls = {}
 
@@ -43,9 +47,6 @@ def test_push_to_hub_tags_uploaded_dataset_revision(tmp_path, monkeypatch):
             return SimpleNamespace(oid="abc123")
 
         def delete_tag(self, repo_id, **kwargs):
-            import requests
-            from huggingface_hub.errors import RevisionNotFoundError
-
             calls["delete_tag"] = {"repo_id": repo_id, **kwargs}
             # Simulate the common case: no stale tag to delete.
             raise RevisionNotFoundError("no such tag", response=requests.Response())
@@ -53,7 +54,12 @@ def test_push_to_hub_tags_uploaded_dataset_revision(tmp_path, monkeypatch):
         def create_tag(self, **kwargs):
             calls["create_tag"] = kwargs
 
-    monkeypatch.setattr("huggingface_hub.HfApi", FakeHfApi)
+    monkeypatch.setattr(lerobot_annotate, "HfApi", FakeHfApi)
+
+    def fake_card_push(self, **kwargs):
+        calls["card_push"] = {"content": str(self), **kwargs}
+
+    monkeypatch.setattr("huggingface_hub.DatasetCard.push_to_hub", fake_card_push)
 
     cfg = SimpleNamespace(
         repo_id="source/dataset",
@@ -62,7 +68,7 @@ def test_push_to_hub_tags_uploaded_dataset_revision(tmp_path, monkeypatch):
         push_commit_message=None,
     )
 
-    _push_to_hub(root, cfg)
+    lerobot_annotate._push_to_hub(root, cfg)
 
     assert calls["create_repo"] == {
         "repo_id": "annotated/dataset",
@@ -71,6 +77,13 @@ def test_push_to_hub_tags_uploaded_dataset_revision(tmp_path, monkeypatch):
         "exist_ok": True,
     }
     assert calls["upload_folder"]["repo_id"] == "annotated/dataset"
+    # The source README must not be copied over: its links (e.g. the
+    # visualize badge) point at the source dataset. A card regenerated for
+    # the target repo is pushed instead.
+    assert "README.md" in calls["upload_folder"]["ignore_patterns"]
+    assert calls["card_push"]["repo_id"] == "annotated/dataset"
+    assert "visualize_dataset?path=annotated/dataset" in calls["card_push"]["content"]
+    assert "source/dataset" not in calls["card_push"]["content"]
     # A stale tag (e.g. from a previous annotation run) is deleted first so
     # the new tag always points at the upload we just made.
     assert calls["delete_tag"] == {

tests/test_rollout.py

@@ -348,3 +348,225 @@ def test_rollout_context_fields():
 
     field_names = {f.name for f in dataclasses.fields(RolloutContext)}
     assert field_names == {"runtime", "hardware", "policy", "processors", "data"}
+
+
+# ---------------------------------------------------------------------------
+# Sync engine: relative-action anchoring (drift-free chunk execution)
+# ---------------------------------------------------------------------------
+
+_REL_ACTION_NAMES = ["j0.pos", "j1.pos", "j2.pos", "gripper.pos"]
+_REL_ACTION_DIM = len(_REL_ACTION_NAMES)
+
+
+def _relative_pre_post(exclude_joints=None):
+    """Pre/post processors wrapping the real relative (caches anchor) and absolute
+    (relative + cached state) steps, mirroring what the sync engine feeds them."""
+    from lerobot.processor import (
+        AbsoluteActionsProcessorStep,
+        RelativeActionsProcessorStep,
+        TransitionKey,
+        create_transition,
+    )
+    from lerobot.utils.constants import OBS_STATE
+
+    relative_step = RelativeActionsProcessorStep(
+        enabled=True, exclude_joints=exclude_joints or [], action_names=list(_REL_ACTION_NAMES)
+    )
+    absolute_step = AbsoluteActionsProcessorStep(enabled=True, relative_step=relative_step)
+
+    class _Pre:
+        steps = [relative_step]
+
+        def __call__(self, observation):
+            # Run the relative step so it caches the anchor, then pass the batch through.
+            transition = create_transition(observation={OBS_STATE: observation[OBS_STATE]})
+            relative_step(transition)
+            return observation
+
+        def reset(self):
+            pass
+
+    class _Post:
+        def __call__(self, action):
+            transition = create_transition(action=action)
+            return absolute_step(transition)[TransitionKey.ACTION]
+
+        def reset(self):
+            pass
+
+    return _Pre(), _Post(), relative_step
+
+
+def _fake_relative_policy(chunk_rel, n_action_steps, chunking=True):
+    """Fake relative-action policy for the sync engine.
+
+    ``chunking=True`` buffers a chunk and serves it one action per tick, calling the
+    public ``predict_action_chunk`` only on refill (pi0/fastwam/lingbot). ``False``
+    returns an action directly and never calls it. The engine's anchor probe keys off
+    that public call, so the fake routes through it rather than any private queue.
+    """
+    from collections import deque
+
+    policy = MagicMock()
+    policy.config.use_amp = False
+    policy.config.action_feature_names = list(_REL_ACTION_NAMES)
+    state = {"predict_calls": 0}
+    queue = deque(maxlen=n_action_steps)
+
+    def predict_action_chunk(_batch=None, **_kwargs):
+        state["predict_calls"] += 1
+        return chunk_rel.unsqueeze(0)  # [B=1, n, dim]
+
+    def select_action(_observation):
+        if not chunking:
+            return chunk_rel[0].unsqueeze(0)
+        if len(queue) == 0:
+            actions = policy.predict_action_chunk(_observation)
+            queue.extend(actions.transpose(0, 1))  # [n, 1, dim]
+        return queue.popleft()
+
+    policy.predict_action_chunk.side_effect = predict_action_chunk
+    policy.select_action.side_effect = select_action
+    policy.reset.side_effect = queue.clear
+    policy._predict_state = state
+    return policy
+
+
+def _build_sync_engine(policy, pre, post):
+    from lerobot.rollout import SyncInferenceEngine
+
+    return SyncInferenceEngine(
+        policy=policy,
+        preprocessor=pre,
+        postprocessor=post,
+        dataset_features={"action": {"names": list(_REL_ACTION_NAMES)}},
+        ordered_action_keys=list(_REL_ACTION_NAMES),
+        task="test",
+        device="cpu",
+        robot_type="mock",
+    )
+
+
+def _obs_frame(state_values):
+    import numpy as np
+
+    return {"observation.state": np.asarray(state_values, dtype=np.float32)}
+
+
+def test_sync_relative_holds_anchor_across_chunk():
+    """Every action popped within a chunk must anchor to the tick-0 state (no drift)."""
+    n = 4
+    # A distinct relative offset per chunk step so a wrong anchor would be visible.
+    chunk_rel = torch.stack([torch.full((_REL_ACTION_DIM,), 0.1 * (i + 1)) for i in range(n)])
+    pre, post, relative_step = _relative_pre_post()
+    policy = _fake_relative_policy(chunk_rel, n_action_steps=n)
+    engine = _build_sync_engine(policy, pre, post)
+
+    assert engine._relative_step is relative_step  # introspection wired the step
+
+    s0 = [1.0, 2.0, 3.0, 4.0]
+    outputs = []
+    for tick in range(n):
+        # Feed a *different* state each tick; a drifting anchor would use it.
+        state = [v + tick for v in s0]
+        outputs.append(engine.get_action(_obs_frame(state)))
+
+    # Exactly one chunk was predicted across the n ticks.
+    assert policy._predict_state["predict_calls"] == 1
+    for tick in range(n):
+        expected = torch.tensor(s0) + chunk_rel[tick]
+        torch.testing.assert_close(outputs[tick], expected)
+
+    # Next tick empties the queue -> fresh chunk -> anchor advances to the new state.
+    s_next = [10.0, 20.0, 30.0, 40.0]
+    out = engine.get_action(_obs_frame(s_next))
+    assert policy._predict_state["predict_calls"] == 2
+    torch.testing.assert_close(out, torch.tensor(s_next) + chunk_rel[0])
+    # The anchor now reflects the fresh-chunk state, not the held one.
+    torch.testing.assert_close(relative_step.get_cached_state(), torch.tensor([s_next]))
+
+
+def test_sync_relative_reset_reanchors_new_episode():
+    """After ``reset()`` the first tick of the next episode anchors to the new state."""
+    n = 3
+    chunk_rel = torch.stack([torch.full((_REL_ACTION_DIM,), 0.2) for _ in range(n)])
+    pre, post, relative_step = _relative_pre_post()
+    policy = _fake_relative_policy(chunk_rel, n_action_steps=n)
+    engine = _build_sync_engine(policy, pre, post)
+
+    # Episode 1: one tick anchors to s0 and leaves cached actions in the queue.
+    engine.get_action(_obs_frame([1.0, 1.0, 1.0, 1.0]))
+    assert policy._predict_state["predict_calls"] == 1
+
+    engine.reset()  # clears the queue and the per-episode chunk flags
+
+    # Episode 2: a fresh state must produce a fresh chunk anchored to that state,
+    # not carry over the previous episode's anchor.
+    s_new = [7.0, 8.0, 9.0, 10.0]
+    out = engine.get_action(_obs_frame(s_new))
+    assert policy._predict_state["predict_calls"] == 2
+    torch.testing.assert_close(out, torch.tensor(s_new) + chunk_rel[0])
+    torch.testing.assert_close(relative_step.get_cached_state(), torch.tensor([s_new]))
+
+
+def test_sync_relative_non_chunking_policy_refreshes_every_tick():
+    """A policy that never calls ``predict_action_chunk`` must not freeze the anchor."""
+    n = 3
+    chunk_rel = torch.stack([torch.full((_REL_ACTION_DIM,), 0.5) for _ in range(n)])
+    pre, post, _ = _relative_pre_post()
+    policy = _fake_relative_policy(chunk_rel, n_action_steps=n, chunking=False)
+    engine = _build_sync_engine(policy, pre, post)
+
+    s0 = [1.0, 1.0, 1.0, 1.0]
+    for tick in range(3):
+        state = [v + tick for v in s0]
+        out = engine.get_action(_obs_frame(state))
+        # Anchor must track the current state every tick (no chunk => no hold).
+        torch.testing.assert_close(out, torch.tensor(state) + chunk_rel[0])
+    assert policy._predict_state["predict_calls"] == 0
+
+
+def test_sync_engine_no_relative_step_is_none():
+    """Without an enabled relative step, the engine takes the plain select_action path."""
+    policy = MagicMock()
+    policy.config.use_amp = False
+    engine = _build_sync_engine(policy, MagicMock(steps=[]), MagicMock())
+    assert engine._relative_step is None
+
+
+def test_sync_relative_exclude_joints_stay_absolute():
+    """With ``exclude_joints``, excluded dims pass through absolute while the relative
+    dims still hold the tick-0 anchor across the chunk."""
+    n = 4
+    # Distinct offset per step *and* per dim so a wrong anchor or a wrong mask shows up.
+    chunk_rel = torch.stack([torch.full((_REL_ACTION_DIM,), 0.1 * (i + 1)) for i in range(n)])
+    pre, post, relative_step = _relative_pre_post(exclude_joints=["gripper"])
+    policy = _fake_relative_policy(chunk_rel, n_action_steps=n)
+    engine = _build_sync_engine(policy, pre, post)
+
+    # gripper (last dim) is kept absolute; j0..j2 are relative.
+    mask = torch.tensor([1.0, 1.0, 1.0, 0.0])
+    s0 = [1.0, 2.0, 3.0, 4.0]
+    outputs = []
+    for tick in range(n):
+        state = [v + tick for v in s0]  # moving state; a drifting anchor would use it
+        outputs.append(engine.get_action(_obs_frame(state)))
+
+    assert policy._predict_state["predict_calls"] == 1  # one chunk held across n ticks
+    for tick in range(n):
+        # relative dims: anchor held at s0; excluded gripper dim: raw predicted value.
+        expected = chunk_rel[tick] + torch.tensor(s0) * mask
+        torch.testing.assert_close(outputs[tick], expected)
+
+
+def test_sync_relative_stop_restores_policy_method():
+    """``stop()`` un-patches the probe so the policy object isn't permanently modified."""
+    n = 3
+    chunk_rel = torch.stack([torch.full((_REL_ACTION_DIM,), 0.2) for _ in range(n)])
+    pre, post, _ = _relative_pre_post()
+    policy = _fake_relative_policy(chunk_rel, n_action_steps=n)
+    original = policy.predict_action_chunk
+    engine = _build_sync_engine(policy, pre, post)
+    assert policy.predict_action_chunk is not original  # probe installed
+    engine.stop()
+    assert policy.predict_action_chunk is original  # restored

tests/utils/test_logging_utils.py

@@ -233,3 +233,37 @@ def test_metrics_tracker_reduce_across_ranks_invokes_reduce():
     # accumulate against the cluster view rather than the stale per-rank sum.
     meter = tracker.update_s
     assert meter.sum / meter.count == pytest.approx(meter.avg)
+
+
+def test_metrics_tracker_update_metrics_registers_and_averages():
+    tracker = MetricsTracker(batch_size=32, num_frames=1000, num_episodes=50, metrics={})
+    tracker.update_metrics({"latent_loss": 0.2, "action_loss": 0.4})
+    tracker.update_metrics({"latent_loss": 0.4, "action_loss": 0.6})
+
+    # New keys are auto-registered as mean-reduced meters and averaged over the window.
+    assert tracker.metrics["latent_loss"].reduction == "mean"
+    assert tracker.metrics["latent_loss"].avg == pytest.approx(0.3)
+    assert tracker.metrics["action_loss"].avg == pytest.approx(0.5)
+    assert tracker.to_dict()["latent_loss"] == pytest.approx(0.3)
+
+
+def test_metrics_tracker_update_metrics_skips_non_numeric():
+    tracker = MetricsTracker(batch_size=32, num_frames=1000, num_episodes=50, metrics={})
+    tracker.update_metrics({"loss": 0.5, "head_mode": "sparse", "enabled": True})
+
+    # strings and bools ignored
+    assert "loss" in tracker.metrics
+    assert "head_mode" not in tracker.metrics
+    assert "enabled" not in tracker.metrics
+
+
+def test_metrics_tracker_update_metrics_does_not_override_caller_meter():
+    # A policy that echoes "loss" in its output dict must not overwrite the caller-owned,
+    # already-aggregated loss meter.
+    metrics = {"loss": AverageMeter("loss", ":.3f", reduction="mean")}
+    tracker = MetricsTracker(batch_size=32, num_frames=1000, num_episodes=50, metrics=metrics)
+    tracker.loss = 1.0  # caller-set optimized loss
+    tracker.update_metrics({"loss": 99.0, "latent_loss": 0.2})
+
+    assert tracker.metrics["loss"].avg == pytest.approx(1.0)  # snapshot ignored
+    assert tracker.metrics["latent_loss"].avg == pytest.approx(0.2)