If two consecutive VLM-emitted subtask spans have ``start`` timestamps
that round to the same source frame after ``snap_to_frame`` (e.g. on
short episodes the VLM sometimes nominates two ~adjacent action
boundaries within one 30 Hz step), the writer emits two
``style=subtask`` rows at the identical persistent timestamp. The
training-time renderer's default binding
``subtask: active_at(t, style=subtask)`` then raises:
ValueError: Ambiguous resolver for style='subtask';
add role=..., tool_name=..., or camera=... to disambiguate.
… and the whole training run dies on the first batch.
Observed concretely on ``pepijn223/super_poulain_vocab2`` (job
22159979): episodes 3 and 30 each had two subtask rows at the same
timestamp (``release yellow cube`` + ``retract arm`` snapping to the
same frame).
Add ``_dedupe_starts_to_distinct_frames`` to walk the cleaned span list
and, whenever a snapped start collides with one already used, push the
later span onto the next free frame timestamp. Both subtasks survive
on distinct timestamps; the renderer can now disambiguate. If the
episode genuinely has no later free frame (extremely unlikely — would
require a same-timestamp collision on the very last frame of the
episode), the later span is dropped with a warning rather than left
to poison the render.
New test ``test_plan_module_bumps_collocated_subtasks_to_distinct_frames``
locks in the contract; full vocabulary suite is 14/14 green.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
The Jaccard-overlap snap was warping VLM output into wrong canonical
labels — e.g. an off-vocab "consult the wizard" span would silently
become "grasp blue cube" if that scored highest. Even with a higher
floor the operator can't tell which subtasks were paraphrases vs
genuine mislabels in the resulting dataset.
Replace with strict exact-match validation + a single targeted retry:
1. Generate subtasks as before.
2. If any returned subtask's normalised form (lowercased, articles
stripped, whitespace collapsed) isn't in the canonical vocab,
fire one retry call naming the offending strings and re-sending
the full canonical list. The retry prompt requires byte-identical
output from the vocab.
3. After the retry, validate again. Spans still off-vocab are
dropped — no fuzzy snapping ever produces a different canonical
label than the VLM actually emitted.
4. If every span ends up off-vocab even after the retry, warn loudly
so the operator extends ``meta/canonical_vocabulary.json`` to
cover the missing phase. The episode is left with empty subtasks
rather than silently fabricated ones — visibility > sweep-under-
the-rug.
Promote ``_NORMALIZE_STRIP_TOKENS`` to a class constant and split the
normalisation helper out so the retry-validation and the final
canonicalisation share one source of truth.
Tests:
- test_plan_module_accepts_article_only_difference: "grasp the blue
cube" still maps to canonical "grasp blue cube" (article-tolerant).
- test_plan_module_retries_when_subtask_off_vocab: paraphrase
triggers the retry which the VLM corrects in pass 2.
- test_plan_module_drops_off_vocab_subtask_after_retry: VLM that
refuses to correct → bad span dropped, in-vocab span kept.
- test_plan_module_empty_when_all_off_vocab_after_retry: every
span off-vocab → episode left empty (no warping).
All 13 vocabulary tests pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
When the canonical vocabulary is enabled and the VLM produces spans
that don't overlap any canonical label, the previous Jaccard-floor
(0.5) dropped them and the episode came out with no subtasks at all
— invisible to the downstream policy. Observed on
``pepijn223/super_poulain_vocab``: some episodes had empty subtask
columns because every VLM-emitted phrase scored below 0.5 against
the discovered vocabulary.
Two-pass canonicalisation:
- First pass keeps the Jaccard floor (lowered from 0.5 → 0.25, to
let mild paraphrases through) and drops everything below.
- If that first pass leaves the episode with **zero** subtasks,
fall back to a second pass that always snaps each VLM span to
its nearest canonical label by Jaccard (no floor). The episode
ends up with subtasks even when the vocabulary missed a phase
— a slightly-wrong canonical label is still closer to the right
motion than nothing at all.
- Log loudly when the fallback fires so the operator can spot
coverage gaps in ``meta/canonical_vocabulary.json``.
- Log a per-episode count at INFO when some (but not all) spans
were dropped so it's visible without spamming the run output.
Promote the Jaccard floor + ignore-tokens to class constants so
they're a single edit point. Add ``force=True`` parameter to
``_canonicalize_subtask`` for the no-floor fallback path.
New test ``test_plan_module_snaps_when_all_off_vocab`` covers the
fallback; existing ``test_plan_module_drops_off_vocab_subtask`` is
adjusted to keep at least one in-vocab span so the floor path can
still fire and is exercised. All 12 vocabulary tests pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
The pipeline previously emitted near-unique subtask + memory phrasings
per episode (free-form LLM rephrasing). On the downstream low-level
policy that collapses the action expert's conditioning to noise: every
episode pairs a different paraphrase with similar motions, so the
expert learns a flat scene-prior that ignores the subtask string —
then at inference the high-level head invents *yet another* paraphrase
and the expert produces tiny "uncertain hover" chunks.
Add a vocabulary-discovery phase (phase 0) that runs once per dataset:
- watches the first ``vocabulary.sample_episodes`` (default 3)
episode videos as one Qwen-VL prompt,
- asks the VLM to derive ~``n_subtask_target`` canonical imperative
subtask labels and ~``n_memory_target`` first-person past-tense
memory milestones that recur across the demos,
- persists them to ``meta/canonical_vocabulary.json`` (human-
inspectable, hand-editable), and
- wires the resulting ``Vocabulary`` into the ``plan`` module so
every per-episode subtask + memory call is constrained to those
exact strings (both as prompt-side instructions *and* post-VLM
validation: paraphrases snap to the closest canonical entry via
token-set overlap; below a 0.5 Jaccard floor the subtask is
dropped rather than warped into something semantically wrong).
Operator workflow:
- first run discovers the vocabulary, writes the JSON, and runs
the ``plan`` module against it,
- subsequent runs reuse the on-disk file (``reuse_existing=True``
default) so hand-edits stick,
- set ``--vocabulary.enabled=False`` to fall back to free-form
generation (the original behaviour).
The discovery prompt forbids gerunds / third-person / adverbs and
caps the lists to the requested counts, matching the Hi-Robot /
π0.6-MEM convention of small per-environment vocabularies. The
``plan`` module's subtask + memory prompts grow a conditional
``{vocabulary_block}`` slot rendered only when a vocabulary is
present; without one the templates collapse to their previous
free-form form.
Tests: 11 new unit tests under tests/annotations/test_vocabulary.py
cover the on-disk round-trip, discovery against the fixture dataset,
``reuse_existing`` short-circuit, paraphrase canonicalisation, off-
vocab subtask dropping, and the no-vocabulary pass-through path.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Port the steerable-pipeline refinements developed on feat/smolvla-on-
steerable back into the annotation pipeline itself:
- module_1_subtasks: imperative verb-first telegraphic labels with a
consistent-object-noun rule and good/bad examples (no hard word cap).
- _generate_plan: drop the VLM round-trip; the plan is now a
deterministic numbered list of still-todo subtasks, re-emitted at
every subtask boundary so it shrinks as work progresses. Removes
module_1_plan.txt.
- VqaConfig.K 3 -> 1: a VQA pair anchors exactly its emission frame, no
stale-label temporal smear.
- lerobot-annotate: tag the pushed dataset with its codebase_version so
LeRobotDataset can resolve a revision and load it.
- module_2_interjection: shorter, more natural mid-task cues.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
PyAV segfaulted (exit 139) decoding the AV1 streams modern LeRobot
datasets use — a SIGSEGV that the per-episode try/except cannot catch,
killing the whole job when the interjections phase started.
Replace the PyAV fallback with _decode_frames_ffmpeg, which shells out
to the ffmpeg CLI: a full ffmpeg build decodes AV1, and a child-process
crash is a catchable non-zero exit rather than a segfault. Decoder chain
is now torchcodec -> ffmpeg. _decode_frames_av stays available behind
video_backend="pyav" for callers that want it.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The pyav fallback routed through lerobot's decode_video_frames(backend=
"pyav"), which uses torchvision.io.VideoReader — removed in torchvision
0.23+. On modern torch stacks (e.g. vllm-openai with torchvision 0.26)
both torchcodec and that path fail, leaving interjection/vqa prompts
without visual context.
Add _decode_frames_av: a self-contained PyAV decoder that picks the
nearest frame per timestamp. It is the always-available tail of the
decoder chain (torchcodec -> pyav) and the target of --video_backend=pyav.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Addresses three of CarolinePascal's frames.py comments (the fourth, the
subprocess re-encode, waits on #3611):
- replace the bespoke _decode_pyav_direct PyAV decoder with
lerobot.datasets.video_utils.decode_video_frames (torchcodec backend,
PyAV fallback) — torchvision's VideoReader removal no longer applies
- frames flow through the provider as torch.Tensor (C, H, W uint8); PIL
is materialised only at the VLM-message boundary in to_image_blocks /
to_video_block, where the chat backends need it
- _decode now returns exactly one frame per timestamp (or [] on failure),
so frames_at pairs them with strict=True
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- name the three modules everywhere (plan / interjections / vqa) instead
of module_1/2/3 — config classes, config fields, executor params,
staging keys and phase names now carry the module name
- rename examples/annotation -> examples/annotations; add the Apache
header to run_hf_job.py
- drop the unused GeneralVqaModule._generate_one
- remove "PR 1" references from comments/docstrings
- frames.py: rely on the always-defined LeRobotDatasetMetadata.camera_keys
- executor.py: read/write meta/info.json via load_info / write_info
- reader.py: load meta/tasks.parquet via io_utils.load_tasks
- make --push_to_hub a bool; push the annotated dataset back to --repo_id
- move the on-disk test dataset builder into tests/fixtures
(build_annotation_dataset); run_e2e_smoke reuses it
- clarify in the docs that the vqa module grounds each pair on a single
frame (K = per-tick anchor count)
- hoist stdlib dynamic imports to module scope
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Both tests were stale relative to design changes that landed earlier on
this branch. Update the tests to match the current production contract.
**``test_module1_attaches_video_block_to_subtask_prompt``**
The test took ``captured[0]`` and asserted on its content blocks, but
Module 1 issues several sub-prompts and the rephrasings call (which is
text-only, no video block) usually lands first. Two fixes:
* The test's intent is "the subtask prompt carries the video block" —
not "the first prompt carries it". Pick the call by content
(``"atomic subtasks"`` keyword in the text block) so the test is
resilient to future reordering of unrelated sub-prompts.
* Set ``n_task_rephrasings=0`` so the rephrasings call is skipped
entirely — keeps the test focused on ``_generate_subtasks``.
**``test_module2_mid_episode_emits_paired_interjection_and_speech``**
Two issues both rooted in design changes on the branch:
1. ``InterjectionsAndSpeechModule._mid_episode_interjections`` now
anchors interjections on subtask boundaries from Module 1's staging
tree, bailing out with zero rows when no spans exist. The production
executor runs Module 1 first; the test ran Module 2 in isolation.
Reproduce the contract by seeding two ``style=subtask`` rows in the
staging before calling Module 2 — gives it the single ``0 → 1``
boundary it needs.
2. The test's stub responder used the marker ``"ONE realistic
interruption"`` to match the interjection prompt, but that string is
from a previous prompt version. The current
``module_2_interjection.txt`` says ``"Write ONE interjection..."`` —
the old prompt asked for counterfactual interjections (e.g. "skip the
wipe"), the new one anchors on the upcoming subtask. Marker updated
to ``"Write ONE interjection"``; canned response wording aligned to
the new design.
Sweep on the language stack: 66 passed, 0 failed (was 64 passed, 2
failed). Pre-commit clean.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
**Critical: video_for_episode was unreachable dead code.**
``video_for_episode`` was indented inside ``_decode_pyav_direct``, after
its ``return`` statement — Python parsed it as a nested function that
never executed. Module 1's ``_episode_video_block`` calls
``self.frame_provider.video_for_episode(record, target_count)`` on the
``use_video_url=False`` path, which would have AttributeError'd on any
real dataset. Tests passed only because they used ``_StubFrameProvider``
/ ``_NullProvider`` which have the method. Moved it to be a proper
method of ``VideoFrameProvider`` (right after ``frames_at``).
**Thread safety on VideoFrameProvider.**
The executor runs Module 1/2/3 phases under a ``ThreadPoolExecutor``, so
the per-instance ``_cache`` dict and the one-shot ``_warned_decode_fail``
flag were exposed to concurrent reads/writes. Added a ``threading.Lock``
field, wrapped cache reads/writes and the warn-flag check-and-set in
``with self._lock:``. Stub fixtures unaffected.
**episode_clip_path is now a method of VideoFrameProvider.**
Used to be a free function reaching into ``provider._meta.episodes`` and
``provider._meta.get_video_file_path`` from outside the class. As a
method it just uses ``self._meta``. The only caller (Module 1) updated;
no external callers.
**Atomic write in LanguageColumnsWriter.**
``pq.write_table(new_table, path)`` was overwriting the parquet shard
in place — a crash mid-write would corrupt the file. Now writes to a
sibling ``.tmp`` and ``Path.replace`` atomically.
**Smaller items:**
* ``executor.py`` docstring opened with "four phases" but listed six.
Now says "six phases" to match.
* ``[annotations]`` extra in ``pyproject.toml`` now includes
``openai>=1.40,<2.0``. Default ``VlmConfig.backend`` is ``"openai"``,
so without it ``_make_openai_client`` would ImportError on a fresh
``uv sync --extra annotations``.
* ``_snap_to_frame`` was duplicated identically in
``plan_subtasks_memory.py`` and ``interjections_and_speech.py``.
Promoted to ``snap_to_frame`` in ``reader.py`` (next to
``EpisodeRecord``); both modules now import it. Backwards-compat alias
not needed — no external callers.
* ``EpisodeRecord.frames_df()`` was re-reading the full parquet on every
call. Now memoizes via a private dataclass field so repeat calls from
different modules pay the cost once. Method signature unchanged.
* ``_extract_first_json_object`` had a redundant ``and not escape`` guard
that was dead because the prior block already handled and reset
``escape``. Replaced with a comment explaining the invariant.
**Pre-existing lint cleanups surfaced once these files entered
pre-commit's scope:**
* dead local ``client = clients[0]`` in ``_make_openai_client`` (the
real round-robin uses ``clients[rr_counter[...]]``).
* ``cmd = ... if "{port}" in cmd else f"...{port}"`` ternary collapse in
``_spawn_parallel_inference_servers``.
* ``seek_pts = 0 if stream.time_base is None else int(...)`` ternary
collapse in ``_decode_pyav_direct``.
* ``# nosec B310`` on the localhost ``urllib.request.urlopen`` probe in
``_server_is_up`` — the URL is the user-configured local-server endpoint
the CLI itself spawned, not arbitrary user input.
**Test added.**
``tests/annotations/test_frames.py`` pins the regression on
``VideoFrameProvider``: asserts ``video_for_episode`` and
``episode_clip_path`` are callable methods (not nested dead code or
free functions), and that the ``_lock`` field is a real
``threading.Lock``.
Sweep: 64 passed, 2 failed (same pre-existing module-impl bugs as
before this commit). Pre-commit clean.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Resolve conflicts and pull in the latest PR 1 fixes.
Conflicts:
- pyproject.toml: PR 1 added `lerobot-rollout` and PR 2 added
`lerobot-annotate` to the same `[project.scripts]` block. Kept both.
- uv.lock: dropped both sides and regenerated against the merged
`pyproject.toml` (PR 2 dropped the `datatrove` dep when distribution
moved to HF Jobs; PR 1's lock didn't have it).
Test follow-up:
- `tests/annotations/test_pipeline_recipe_render.py` — PR 1 deleted
`src/lerobot/configs/recipes/pi05_hirobot.yaml` (review feedback:
remove the canonical-recipe file; recipes are user-supplied). The
cross-PR contract this test guards is "the recipe DSL renders
non-empty messages from pipeline output", which doesn't depend on
any specific YAML, so the test now builds an inline blend recipe
with the same coverage. Passes.
Sweep: 82 passed, 2 failed (pre-existing module-impl bugs:
`test_module1_attaches_video_block_to_subtask_prompt`,
`test_module2_mid_episode_emits_paired_interjection_and_speech`).
The PR 1 carryover (`test_emitted_at_raises_on_ambiguous_per_camera_vqa`)
is now passing — the merge brought in PR 1's tightened `_select_one`
ambiguity check.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat: add pretrained vision encoder weights for diffusion and vqbet
* fix test by re-generating artifacts
---------
Co-authored-by: Steven Palma <imstevenpmwork@ieee.org>
* **Float tolerance in `emitted_at` for persistent styles.** The
``_timestamp(row) == t`` exact-equality check silently missed any
caller that derived ``t`` arithmetically (e.g. ``frame_idx / fps``)
even though the parquet timestamp would only differ by ULPs. Added
``EMITTED_AT_TOLERANCE_S = 0.1`` and check ``abs(...) <= tolerance``
instead, with a docstring explaining why exact equality wasn't
enough and why 0.1 s is safe at typical 30–100 Hz control rates.
Test asserts the new behavior at half-window (matches) and
double-window (no match) using the constant so it stays in sync.
* **`MessageTurn.stream` is required at construction.** It was typed
``MessageStream | None = None`` so YAML could omit ``stream:`` and
pass the dataclass invariant — but ``_validate_rendered`` rejected
``None`` streams later, surfacing the error at the first sample
instead of at recipe load. Now ``__post_init__`` raises
``ValueError`` if ``stream`` is ``None``, with the list of valid
streams in the message. The redundant late-stage check in
``_validate_rendered`` is replaced with a one-line comment that
cites the upstream invariant. Test pins the new construction-time
rejection.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* **#2 — dedupe `_PLACEHOLDER_RE`.** The same regex was compiled in
`recipe.py` and `language_render.py`. Promote to module-level
`PLACEHOLDER_RE` in `recipe.py` (its primary owner — declares
template syntax) and import from `language_render.py`.
* **#3 — centralize language column names.** `io_utils.py` had
hardcoded `{"language_persistent", "language_events"}` literals at
two sites. Replace with `LANGUAGE_COLUMNS` import so a future column
rename can't silently desync.
* **#4 — defensive collate preserved-keys.** `lerobot_collate_fn`
silently filtered language fields from samples that didn't have
them, which would hand downstream consumers a preserved list
shorter than the tensor batch. Now: if any sample carries a key,
every sample in the batch must carry it; otherwise raise a
`ValueError` so the upstream rendering bug surfaces at the boundary.
* **#5 — `_scalar` rejects non-singleton lists.** Previously a zero-
or multi-element list fell through and triggered confusing
`float([])` errors downstream. Now raises `ValueError` with the
actual length.
* **#6 — refactor `_extract_complementary_data`.** Replace 11 lines
of `key = {... if ... else {}}` plus an 11-line splat dict with a
single `_COMPLEMENTARY_KEYS` tuple iterated once.
* **#7 — document `EXTENDED_STYLES`.** Was an empty `set()` with no
comment. Add a docstring explaining it's an intentional extension
point: downstream modules append project-local styles before
`column_for_style` is called.
* **#9 — `tools.mdx` notes the runtime layer is future work.** The
page referenced `src/lerobot/tools/`, `registry.py`, and
`get_tools(meta)` — none exist in this PR. Added a callout at the
start of "How to add your own tool" plus a note on the
implementations paragraph.
* **#10 — tests for YAML round-trip, malformed rows, blend
validation.** `test_recipe.py` grew from 1 case to 12 covering:
blend-or-messages exclusivity, target-turn requirement, blend
emptiness, weight presence/positivity, nested-blend rejection,
`from_dict` with nested blends, `from_yaml` / `load_recipe`
agreement, top-level non-mapping rejection. Added a malformed-row
test for `_normalize_rows` that asserts non-dict entries raise
`TypeError`.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(policies): add EO-1 model
* chore(eo1): adjust policy_eo1_README.md to to avoid duplicate with eo1.mdx
* chore(eo1): remove policy_eo1_README.md, link eo1.mdx in policy folder
---------
Co-authored-by: Pepijn <138571049+pkooij@users.noreply.github.com>
* **`meta.tools` actually reads `info.json["tools"]`.** `DatasetInfo`
had no `tools` field, so `from_dict` silently dropped the key (it
warned about unknown fields then discarded them) and the property
always returned `DEFAULT_TOOLS`. Added `tools: list[dict] | None`
to the dataclass; `to_dict()` drops it when unset so existing
datasets keep a clean `info.json`. Fixed the accessor to read
`self.info.tools` (the previous `.get(...)` would have raised
AttributeError on the dataclass anyway). Added regression tests:
fallback when absent, round-trip from disk, and round-trip
through `DatasetInfo.from_dict` / `to_dict`.
* **`motion` is not view-dependent — fix the docs.** The mdx claimed
rows of style `motion` must carry `camera`, but `VIEW_DEPENDENT_STYLES
= {"vqa", "trace"}` and the validator agrees: motion primitives are
joint/Cartesian-frame, not pixel-space. Updated both call-out
paragraphs in `language_and_recipes.mdx`.
* **Conditional `collate_fn` swap.** Added `meta.has_language_columns`
and gate the `lerobot_collate_fn` swap in `lerobot_train.py` on it,
so non-language datasets keep PyTorch's `default_collate`. Also
added a pass-through test in `test_collate.py` that asserts on a
plain tensor batch the custom collate matches `default_collate`
key-for-key, plus a test for the `None`-sample drop path.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
`lerobot.processor` re-exported `RenderMessagesStep` at the package
level, so importing anything from `lerobot.processor` pulled in
`lerobot.datasets.language` → `lerobot.datasets/__init__.py` →
`require_package("datasets")`, which fails in the Tier 1 base install
that intentionally omits the `[dataset]` extra. The chain bricked
collection for unrelated suites (`tests/policies/pi0_pi05/...`,
`tests/envs/...`, etc.).
* Stop re-exporting `RenderMessagesStep` from `lerobot.processor`. The
only consumer (the test) already imports from the submodule.
Document the deliberate omission in the module docstring.
* Add `pytest.importorskip("datasets", ...)` (and `pandas` where
needed) at the top of the four PR-added tests that exercise the
language stack:
- tests/datasets/test_language.py
- tests/datasets/test_language_render.py
- tests/processor/test_render_messages_processor.py
- tests/utils/test_collate.py
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Drop the unused `events` kwarg from `active_at`/`nth_prev`/`nth_next`;
only `emitted_at` actually consults events. The dispatcher in
`_resolve_spec` now passes events conditionally.
* Replace the dual `_persistent_sort_key`/`_event_sort_key` pair with a
single `_row_sort_key` and drop the `sort_key` parameter from
`_select_one`. Event rows lack `timestamp` (it is implicit in the
frame) and now default to `0.0` for sort purposes — the
`(style, role)` tiebreaker is unchanged.
* Inline `_select_latest` into `active_at` (its only caller).
* Collapse `emitted_at`'s dual-branch into one `_select_one` call.
* Tighten `_validate_persistent_resolver` to a single
`column_for_style(style) != LANGUAGE_PERSISTENT` check.
* Parameterize `test_per_camera_blend_renders_both_views` over the two
cameras and factor the sub-recipe builder into `_vqa_subrecipe` so
the test no longer hand-rolls two near-identical recipe blocks.
Net -98 LOC; behavior, public resolver names, and test expectations
unchanged.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Ensure annotated datasets advertise language columns in meta/info.json so non-streaming dataset loads cast against the rewritten parquet schema.
Co-authored-by: Cursor <cursoragent@cursor.com>
PR 2 used to write a top-level ``tools`` column on every parquet shard
holding the JSON schema for the ``say`` tool, broadcast identically
across every row. That extends PR 1's schema for no real information
gain — the schema is a fixed code constant, parquet's RLE/dict encoding
collapses it on disk anyway, and HF/TRL chat-template consumers can
just import the constant directly.
PR 2 should fill in PR 1's existing schema, not add to it. So:
- ``writer.py``: stop emitting the ``tools`` column. Strip any legacy
``tools`` column from older shards on rerun so the schema converges to
v3.1. ``SAY_TOOL_SCHEMA`` stays as a public constant (now joined by
``DEFAULT_TOOLS = [SAY_TOOL_SCHEMA]``); chat-template policies and the
visualizer import them directly.
- ``test_writer.py``: replace the "tools column present" assertion with
one that explicitly checks the column is absent, plus a new test
asserting the constant's shape.
- ``test_pipeline_recipe_render.py``: drop the tools-column read; assert
it's not present in the rewritten parquet.
- ``annotation_pipeline.mdx``: update the writer description to note the
parquet stays small and the schema lives as a code constant.
If multi-tool-set support ever becomes real (datasets with different
tool inventories), the right home is ``meta/info.json["tools"]`` —
adding it later is non-breaking; ripping out a parquet column already
shipped is not.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Module 3 now produces one (vqa, user) + (vqa, assistant) pair per
emission tick *per camera* rather than only against the dataset's first
camera. Each emitted row carries the `camera` field added in PR 1
(language-columns), so the resolver can disambiguate per-camera VQA via
`emitted_at(t, style=vqa, role=assistant, camera=...)` without ambiguity.
- `frames.py`: `FrameProvider` Protocol gains a `camera_keys` property
and a `camera_key=` argument on `frames_at` / `video_for_episode`.
`VideoFrameProvider` exposes every `observation.images.*` key the
dataset declares (not just the first) and keys its decode cache on
`(episode, camera, timestamp)` so per-camera reads don't collide.
Module 1 / 2 keep their old single-camera behaviour by leaving
`camera_key=None` (falls back to the default camera).
- `modules/general_vqa.py`: `run_episode` iterates `frame_provider
.camera_keys` for each emission tick, builds one prompt per camera,
batches all of them through the VLM, and stamps the resulting rows
with `camera=<that key>`. Empty `camera_keys` (null provider) makes
the module a no-op rather than silently emitting untagged rows.
- `writer.py`: `_normalize_persistent_row` / `_normalize_event_row`
carry `camera` through and call `validate_camera_field` so the
invariant is enforced at the writer boundary. Event sort key now
includes `camera` for deterministic ordering when several cameras
share `(timestamp, style, role)`. `speech_atom` sets `camera=None`.
- `validator.py`: `StagingValidator` gains a `dataset_camera_keys`
field; `_check_camera_field` enforces the invariant and cross-checks
every view-dependent row's `camera` against the dataset's known video
keys. New `_check_vqa_uniqueness_per_frame_camera` flags duplicate
`(vqa, role)` pairs at the same `(t, camera)`.
- `lerobot_annotate.py`: passes the live frame provider's
`camera_keys` into the validator so the cross-check uses the actual
dataset camera set.
- Tests: `_StubFrameProvider` exposes `camera_keys` and accepts the new
`camera_key=` kwarg. `test_module3_vqa_unique_per_frame_and_camera`
configures two cameras and asserts both are represented, that every
emitted row has a `camera` tag, and that uniqueness holds per
`(timestamp, camera, role)`.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replaces keyframe sampling with a single Qwen-VL video block covering
the whole demonstration. The model pools temporally itself and chooses
where to cut subtasks — no stride, no count, no keyframe count knob to
tune.
- frames.py: ``FrameProvider`` gains ``video_for_episode(record,
max_frames)``; ``VideoFrameProvider`` samples up to ``max_frames``
uniformly across the episode duration; ``_NullProvider`` returns []
for the no-video fallback. New ``to_video_block`` helper.
- Module 1: drops keyframe sampling. The subtask prompt now goes out as
``[{"type":"video", "video":[<frames>]}, {"type":"text", ...}]`` and
the prompt template asks the model to "watch the whole clip, then
segment it" with cut points decided from gripper/contact/regrasp
events the model sees.
- Module1Config: ``keyframes_per_episode`` removed; replaced with
``max_video_frames: int = 32`` (model-capacity bound, not annotation
logic).
- Test: ``test_module1_attaches_video_block_to_subtask_prompt`` locks in
the single-video-block invariant.
- Stub-VLM markers updated: tests now key on "atomic subtasks" instead
of the old "Decompose the demonstration" phrase that no longer
appears in the prompt.
- Docs: updated to describe the whole-episode video-block behavior and
the no-video fallback.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Closes the visual-grounding gap flagged after the initial PR review:
modules now decode actual camera frames at the relevant timestamps and
attach them as `{"type":"image", "image":<PIL>}` content blocks to the
VLM prompts.
- New `frames.py`:
- `FrameProvider` Protocol; `VideoFrameProvider` decodes from the
dataset's first `observation.images.*` stream via
`LeRobotDatasetMetadata.get_video_file_path` and
`decode_video_frames`, with the same `from_timestamp` shift the main
dataset uses.
- Per-process LRU cache so co-timestamped Module 1 plan-update + Module
2 calls share decode work.
- `make_frame_provider` falls back to a null provider when the dataset
has no video tracks → text-only prompts (graceful absence).
- Modules 1/2/3 take an optional `frame_provider` (default null) and
prepend image blocks before the text block.
- Module 1 attaches `keyframes_per_episode` keyframes to the subtask
decomposition prompt.
- Module 2 attaches the frame at the interjection timestamp.
- Module 3 attaches the exact emission frame to each VQA pair.
- VlmConfig: backend now defaults to `vllm`; default model is
`Qwen/Qwen3.6-27B-FP8`. New knobs: `--vlm.tensor_parallel_size`,
`--vlm.camera_key` (override the keyframe stream).
- `_make_vllm_client` honours `tensor_parallel_size` so 27B-FP8 sharded
on 2× GPUs works out of the box.
- `test_module3_attaches_frame_image_block_to_prompt` asserts modules
emit one image block per VQA prompt at the exact emission timestamp.
- Docs: example switched to `imstevenpmwork/super_poulain_draft` +
Qwen3.6-27B-FP8 + tensor_parallel_size=2; documents the keyframe
attachment behaviour and the no-video fallback.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds task-prompt diversity (Xiao 2022 / CAST) without touching
``meta/tasks.parquet`` or forcing recipes to opt in. The plan reserved
``task_aug`` as a future style; this lands it now.
- ``language.py``: add ``task_aug`` to ``CORE_STYLES`` and
``PERSISTENT_STYLES``. ``column_for_style("task_aug")`` returns
``language_persistent`` so PR 2 writers route it correctly.
- ``language_render.py``: ``_resolve_task`` now consults the persistent
slice for rows of ``style="task_aug", role="user"``. When any exist
it picks one deterministically by ``sample_idx`` (blake2b-keyed, not
Python's randomized hash) so an epoch sees every rephrasing of every
episode while the same sample still resolves identically across
reruns. Falls back to the canonical ``meta/tasks.parquet`` task when
no rephrasings are present, so existing datasets and unannotated runs
keep their behaviour. Explicit ``task=`` overrides still win.
- Tests: rephrasing coverage across samples, determinism on repeat
``sample_idx``, fallback when persistent has no ``task_aug`` rows,
and explicit override priority.
Recipes get this for free: any ``${task}`` placeholder rotates through
the available rephrasings. Recipes that want the literal canonical task
can override the binding.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Motion primitives are described in robot-frame (joint / Cartesian) terms,
not pixel space, so they are camera-agnostic. Only `vqa` (event) and
`trace` (event, pixel-trajectory) are view-dependent.
The `camera` field stays on PERSISTENT_ROW_FIELDS for schema symmetry —
the validator, resolver, and HF feature mapping behave identically across
the two columns regardless of which styles populate `camera` today —
but persistent rows now always have `camera=None` in practice.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds a nullable `camera` field to the language row struct (both persistent
and event variants) so view-dependent styles like `vqa` can carry which
`observation.images.*` view they were grounded against. Without this,
multi-camera datasets ended up with multiple `(vqa, role)` rows at the
same timestamp that the resolver could not disambiguate.
- `language.py`: add `camera` to PERSISTENT_ROW_FIELDS / EVENT_ROW_FIELDS,
to both Arrow struct types and the HF datasets feature mappings;
introduce VIEW_DEPENDENT_STYLES = {vqa, motion, trace} plus
`is_view_dependent_style` and `validate_camera_field` helpers (camera
required iff style is view-dependent).
- `language_render.py`: thread an optional `camera=` kwarg through every
resolver (`active_at`, `emitted_at`, `nth_prev`, `nth_next`) and through
`_matching_rows` / `_select_*`, so recipes can disambiguate per-camera
VQA with `emitted_at(t, style=vqa, role=assistant, camera=...)`.
Without a `camera` filter, multi-row matches keep raising the existing
ambiguity error — which is the desired behaviour on multi-camera data.
- `recipes/pi05_hirobot.yaml`: replace the single `ask_vqa` branch with
`ask_vqa_top` and `ask_vqa_wrist` per-camera sub-recipes (each carrying
the matching image block), keeping the original 0.20 budget and
documenting the customization point for datasets with different cameras.
- Tests: schema test asserts the new field order; new tests cover
`is_view_dependent_style`, `validate_camera_field` (both required and
forbidden directions), per-camera `emitted_at` filtering, and the
ambiguity error when two cameras emit `(vqa, assistant)` at the same
timestamp without a `camera=` filter. RenderMessagesStep + dataset
passthrough fixtures updated to include the new field.
- `docs/source/language_and_recipes.mdx`: document the `camera` field,
the per-camera resolver pattern, and the canonical recipe convention.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* refactor(datasets): replace untyped dict with typed DatasetInfo dataclass
Introduce typed DatasetInfo dataclass to replace untyped dict representation of info.json.
Changes:
- Add DatasetInfo dataclass with explicit fields and validation
- Implement __post_init__ for shape conversion (list ↔ tuple)
- Add dict-style compatibility layer (__getitem__, __setitem__, .get())
- Add from_dict() and to_dict() for JSON serialization
- Update io_utils to use load_info/write_info with DatasetInfo
- Update dataset utilities and metadata to use attribute access
- Remove aggregate.py dict-style field access
- Add tests fixture support for DatasetInfo
Benefits:
- Type safety with IDE auto-completion
- Validation at construction time
- Explicit schema documentation
* fix pre-commit
* update docstring inside DatasetInfo.from_dict()
* sorts the unknown to have deterministic output
Signed-off-by: Maxime Ellerbach <maxime@ellerbach.net>
* refactoring the last few old fieds
* fix crop dataset roi type mismatch
* use consistantly int for data and video_files_size_in_mb
---------
Signed-off-by: Maxime Ellerbach <maxime@ellerbach.net>
Co-authored-by: jjolla93 <jjolla93@gmail.com>
* feat(rewards): add RewardModelConfig and PreTrainedRewardModel base classes
* refactor(rewards): migrate Classifier from policies/sac/reward_model/ to rewards/classifier/
* refactor(rewards): migrate SARM from policies/sarm/ to rewards/sarm/
* refactor(rewards): add rewards/factory.py and remove reward model code from policies/factory.py
* refactor(rewards): update imports and delete old reward model locations
* test(rewards): add reward model tests and update existing test imports
* fix(rewards): restore full Classifier and SARM implementations
* test(rewards): restore missing CUDA and mixed precision classifier processor tests
* refactor(lerobot_train.py): remove rabc specific configuration and replace it with a generic samplerweight class in lerobot_train
* refactor(lerobot_train.py): add missing sampling weight script
* linter + missing files
* add testing for sampl weighter
* revert some useless changes, improve typing
* update docs
* add automatic detection of the progress path
* remove type exp
* improve comment
* fix: move rabc.py to rewards/sarm/ and update import paths
* refactor(imports): update reward model imports to new module structure
* refactor(imports): update reward model imports to reflect new module structure
* refactor(imports): conditionally import pandas based on availability
* feat(configs): add reward_model field to TrainPipelineConfig and Hub fields to RewardModelConfig
* refactor(policies): remove reward model branches from policy factory and __init__
* refactor(rewards): expand __init__ facade and fix SARMConfig __post_init__ crash
* feat(train): route reward model training through rewards/factory instead of policies/factory
* refactor(train): streamline reward model training logic
* fix(rewards): ensure FileNotFoundError is raised for missing config_file
* refactor(train): update __get_path_fields__ to include reward_model for config loading
* refactor(classifier): remove redundant input normalization in predict_reward method
* fix(train): raise ValueError for non-trainable reward models in train function
* refactor(pretrained_rm): add model card template
* refactor(tests): reward models
* refactor(sarm): update reset method and remove unused action prediction methods
* refactor(wandb): differentiate tags for reward model and policy training in cfg_to_group function
* fix(train): raise ValueError for PEFT usage in reward model training
* refactor(rewards): enhance RewardModelConfig with device handling and delta indices properties
---------
Co-authored-by: Michel Aractingi <michel.aractingi@huggingface.co>
Promote the previously-reserved motion/trace styles to first-class core
styles. motion routes to language_persistent (it tracks robot state over
time); trace routes to language_events (single-moment annotations).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(envs): add RoboMME benchmark integration
- RoboMME env wrapper with image/wrist_image/state observations
- Docker image with Vulkan, SAPIEN, mani-skill deps
- CI workflow: 1-episode smoke eval with pepijn223/smolvla_robomme
- preprocess_observation: handle image/wrist_image/state keys
- pyproject.toml: robomme extra
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* refactor(docker): rebase RoboMME image on huggingface/lerobot-gpu
Mirror the libero/metaworld pattern: start from the nightly GPU image
(which already has apt deps, uv, venv, and lerobot[all] preinstalled)
and only layer on what RoboMME uniquely needs — the Vulkan libs
ManiSkill/SAPIEN requires, plus the robomme extra with the
gymnasium/numpy overrides.
Drops 48 lines of duplicated base setup (CUDA FROM, python install,
user creation, venv init, base apt deps) that the nightly image already
provides. Net: 102 → 54 lines.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* docs(robomme): drop prototype-branch note and move dataset to lerobot/robomme
- Remove the "Related work" block referencing the prototype branch
feat/robomme-integration; the PR stands on its own.
- Point all dataset references at lerobot/robomme (docs, env module
docstring, RoboMMEEnvConfig docstring) — this is the canonical HF
location once the dataset is mirrored.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(robomme): make docs build + fast tests green
1. Docs: add robomme to _toctree.yml under Benchmarks so doc-builder's
TOC integrity check stops rejecting the new page.
2. Fast tests: robomme's mani-skill transitively pins numpy<2 which is
unsatisfiable against the project's numpy>=2 base pin, so `uv sync`
couldn't resolve a universal lockfile.
Drop robomme as a pyproject extra entirely — it truly cannot coexist
with the rest of the dep tree. The Dockerfile installs robomme
directly from its git URL via `uv pip install --override`, which was
already the runtime path. pyproject, docs, env docstrings, and the
CI job comment all now point to the docker-only install.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* test(robomme): realign unit tests with current env API
The tests were written against an earlier env layout and never updated when
the wrapper was refactored, so CI's fast-test job was failing with:
- KeyError: 'front_rgb' / 'wrist_rgb' — these were renamed to the
lerobot-canonical 'image' / 'wrist_image' keys (matching the dataset
columns and preprocess_observation's built-in fallbacks).
- AssertionError: 'robomme' not in result — create_robomme_envs now
returns {task_name: {task_id: env}}, not {'robomme': {...}}, so
comma-separated task lists work.
- ModuleNotFoundError: lerobot.envs.lazy_vec_env — LazyVectorEnv was
removed; create_robomme_envs is straightforward synchronous now.
Rewrite the 7 failing cases against the current API, drop the three
LazyVectorEnv tests, and add a multi-task test so the new comma-separated
task parsing is covered. Stub install/teardown is moved into helpers
(`_install_robomme_stub` / `_uninstall_robomme_stub`) so individual tests
stop repeating six boilerplate lines.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* ci: point benchmark eval checkpoints at the lerobot/ org mirrors
pepijn223/smolvla_* → lerobot/smolvla_* across every benchmark job in
this branch (libero, metaworld, and the per-branch benchmark). The
checkpoints were mirrored into the lerobot/ org and that's the canonical
location going forward.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: integrate PR #3311 review feedback
- envs: rename obs keys to pixels/image, pixels/wrist_image, agent_pos
- envs: add __post_init__ for dynamic action_dim in RoboMMEEnv config
- envs: remove special-case obs conversion in utils.py (no longer needed)
- ci: add Docker Hub login, HF_USER_TOKEN guard, --env.task_ids=[0]
- scripts: extract_task_descriptions supports multiple task_ids
- docs: title to # RoboMME, add image, restructure eval section
- tests: update all key assertions to match new obs naming
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(docs): use correct RoboMME teaser image URL
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* ci(robomme): smoke-eval 10 tasks instead of 5
Broader coverage on the RoboMME benchmark CI job: bump the smoke eval
from 5 tasks to 10 (one episode each), all drawn from ROBOMME_TASKS.
Tasks now run: PickXtimes, BinFill, StopCube, MoveCube, InsertPeg,
SwingXtimes, VideoUnmask, ButtonUnmask, PickHighlight, PatternLock.
Updated the parse_eval_metrics.py `--task` label from the single
`PickXtimes` stub to the full comma list so the metrics artifact
reflects what was actually run. `parse_eval_metrics.py` already reads
`overall` for multi-task runs, so no parser change is needed.
Made-with: Cursor
* fix(robomme): nest `pixels` as a dict so preprocess_observation picks it up
`_convert_obs` was returning flat keys (`pixels/image`,
`pixels/wrist_image`). `preprocess_observation()` in envs/utils.py
keys off the top-level `"pixels"` entry and, not finding it,
silently dropped every image from the batch. The policy then saw
zero image features and raised
ValueError: All image features are missing from the batch.
Match the LIBERO layout: return
`{"pixels": {"image": ..., "wrist_image": ...}, "agent_pos": ...}`
and declare the same shape in `observation_space`.
Made-with: Cursor
* fix(robomme): align docs and tests with nested pixels obs layout
Addresses PR #3311 review feedback:
- Docs: correct observation keys to `pixels/image` / `pixels/wrist_image`
(mapped to `observation.images.image` / `observation.images.wrist_image`)
and drop the now-obsolete column-rename snippet.
- Tests: assert `result["pixels"]["image"]` instead of flat `pixels/image`,
matching the nested layout required by `preprocess_observation()`.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(envs): preserve AsyncVectorEnv metadata/unwrapped in lazy eval envs
Port of #3416 onto this branch.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* ci: gate Docker Hub login on secret availability
Fork PRs cannot access `secrets.DOCKERHUB_LEROBOT_{USERNAME,PASSWORD}`,
which made every benchmark job fail at the login step. Gate the login
on the env-var expansion of the username so the step is skipped (not
failed) when secrets are absent.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(robomme): address review feedback
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat(envs): add RoboTwin 2.0 benchmark integration
- RoboTwinEnvConfig with 4-camera setup (head/front/left_wrist/right_wrist)
- Docker image with SAPIEN, mplib, CuRobo, pytorch3d (Python 3.12)
- CI workflow: 1-episode smoke eval with pepijn223/smolvla_robotwin
- RoboTwinProcessorStep for state float32 casting
- Camera rename_map: head_camera/front_camera/left_wrist -> camera1/2/3
* fix(robotwin): re-enable autograd for CuRobo planner warmup and take_action
lerobot_eval wraps the full rollout in torch.no_grad() (lerobot_eval.py:566),
but RoboTwin's setup_demo → load_robot → CuroboPlanner(...) runs
motion_gen.warmup(), which invokes Newton's-method trajectory optimization.
That optimizer calls cost.backward() internally, which raises
RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn
when autograd is disabled. take_action() hits the same planner path at every
step. Wrap both setup_demo and take_action in torch.enable_grad() so CuRobo's
optimizer can build its computation graph. Policy inference is unaffected —
rollout()'s inner torch.inference_mode() block around select_action() is
untouched, so we still don't allocate grad buffers during policy forward.
* fix(robotwin): read nested get_obs() output and use aloha-agilex camera names
RoboTwin's base_task.get_obs() returns a nested dict:
{"observation": {cam: {"rgb": ..., "intrinsic_matrix": ...}},
"joint_action": {"left_arm": ..., "left_gripper": ...,
"right_arm": ..., "right_gripper": ...,
"vector": np.ndarray},
"endpose": {...}}
Our _get_obs was reading raw["{cam}_rgb"] / raw["{cam}"] and raw["joint_action"]
as if they were flat, so np.asarray(raw["joint_action"], dtype=float64) tripped
on a dict and raised
TypeError: float() argument must be a string or a real number, not 'dict'
Fix:
- Pull images from raw["observation"][cam]["rgb"]
- Pull joint state from raw["joint_action"]["vector"] (the flat array)
- Update the default camera tuple to (head_camera, left_camera, right_camera)
to match RoboTwin's actual wrist-camera names (envs/camera/camera.py:135-151)
* refactor(robotwin): drop defensive dict guards, cache black fallback frame
_get_obs was guarding every dict access with isinstance(..., dict) in case
RoboTwin's get_obs returned something else — but the API contract
(envs/_base_task.py:437) always returns a dict, so the guards were silently
masking real failures behind plausible-looking zero observations. Drop them.
Also:
- Cache a single black fallback frame in __init__ instead of allocating
a fresh np.zeros((H, W, 3), uint8) for every missing camera on every
step — the "camera not exposed" set is static per env.
- Only allocate the zero joint_state on the fallback path (not unconditionally
before the real value overwrites it).
- Replace .flatten() with .ravel() (no copy when already 1-D).
- Fold the nested-dict schema comment and two identical torch.enable_grad()
rationales into a single Autograd section in the class docstring.
- Fix stale `left_wrist` camera name in the observation docstring.
* fix(robotwin): align observation_space dims with D435 camera output
lerobot_eval crashed in gym.vector's SyncVectorEnv.reset with:
ValueError: Output array is the wrong shape
because RoboTwinEnvConfig declared observation_space = (480, 640, 3) but
task_config/demo_clean.yml specifies head_camera_type=D435, which renders
(240, 320, 3). gym.vector.concatenate pre-allocates a buffer from the
declared space, so the first np.stack raises on shape mismatch.
Changes:
- Config defaults now 240×320 (the D435 dims in _camera_config.yml), with
a comment pointing at the source of truth.
- RoboTwinEnv.__init__ accepts observation_height/width as Optional and
falls back to setup_kwargs["head_camera_h/w"] so the env is self-consistent
even if the config is not in sync.
- Config camera_names / features_map use the actual aloha-agilex camera
names (head_camera, left_camera, right_camera). Drops the stale
"front_camera" and "left_wrist"/"right_wrist" entries that never matched
anything RoboTwin exposes.
- CI workflow's rename_map updated to match the new camera names.
* fix(robotwin): expose _max_episode_steps for lerobot_eval.rollout
rollout() does `env.call("_max_episode_steps")` (lerobot_eval.py:157) to
know when to stop stepping. LiberoEnv and MetaworldEnv set this attribute;
RoboTwinEnv was tracking the limit under `episode_length` only, so the call
raised AttributeError once CuRobo finished warming up.
* fix(robotwin): install av-dep so lerobot_eval can write rollout MP4s
write_video (utils/io_utils.py:53) lazily imports PyAV via require_package
and raises silently inside the video-writing thread when the extra is not
installed — so the eval itself succeeds with pc_success=100 but no MP4
ever lands in videos/, and the artifact upload reports "No files were
found". Add av-dep to the install line (same pattern as the RoboMME image).
* feat(robotwin): eval 5 diverse tasks per CI run with NL descriptions
Widen the smoke eval from a single task (beat_block_hammer) to five:
click_bell, handover_block, open_laptop, stack_blocks_two on top of the
original. Each gets its own rollout video in videos/<task>_0/ so the
dashboard can surface visually distinct behaviours.
extract_task_descriptions.py now has a RoboTwin branch that reads
`description/task_instruction/<task>.json` (already shipped in the clone
at /opt/robotwin) and pulls the `full_description` field. CI cds into
the clone before invoking the script so the relative path resolves.
parse_eval_metrics.py is invoked with the same 5-task list so the
metrics.json embeds one entry per task.
* ci: point benchmark eval checkpoints at the lerobot/ org mirrors
pepijn223/smolvla_* → lerobot/smolvla_* across every benchmark job in
this branch (libero, metaworld, and the per-branch benchmark). The
checkpoints were mirrored into the lerobot/ org and that's the canonical
location going forward.
* refactor(robotwin): rebase docker image on huggingface/lerobot-gpu
Mirror the libero/metaworld/libero_plus/robomme pattern: start from the
nightly GPU image (apt deps, python, uv, venv, lerobot[all] already
there) and layer on only what RoboTwin 2.0 uniquely needs —
cuda-nvcc + cuda-cudart-dev (CuRobo builds from source), Vulkan libs +
NVIDIA ICD (SAPIEN renderer), sapien/mplib/open3d/pytorch3d/curobo
installs, the mplib + sapien upstream patches, and the TianxingChen
asset download.
Drops ~90 lines of duplicated base setup (CUDA FROM, apt python, uv
install, user creation, venv init, base lerobot install). 199 → 110.
Also repoint the docs + env docstring dataset link from
hxma/RoboTwin-LeRobot-v3.0 to the canonical lerobot/robotwin_unified.
* docs(robotwin): add robotwin to _toctree.yml under Benchmarks
doc-builder's TOC integrity check was rejecting the branch because
docs/source/robotwin.mdx existed but wasn't listed in _toctree.yml.
* fix(robotwin): defer YAML lookup and realign tests with current API
__init__ was eagerly calling _load_robotwin_setup_kwargs just to read
head_camera_h/w from the YAML. That import (`from envs import CONFIGS_PATH`)
required a real RoboTwin install, so constructing the env — and thus every
test in tests/envs/test_robotwin.py — blew up with ModuleNotFoundError
on fast-tests where RoboTwin isn't installed.
Replace the eager lookup with DEFAULT_CAMERA_H/W constants (240×320, the
D435 dims baked into task_config/demo_clean.yml). reset() still resolves
the full setup_kwargs lazily — that's fine because reset() is only
called inside the benchmark Docker image where RoboTwin is present.
Also resync the test file with the current env API:
- mock get_obs() as the real nested {"observation": {cam: {"rgb": …}},
"joint_action": {"vector": …}} shape
- patch both _load_robotwin_task and _load_robotwin_setup_kwargs
(_patch_load → _patch_runtime)
- drop `front_camera` / `left_wrist` from assertions — aloha-agilex
exposes head_camera + left_camera + right_camera, not those
- black-frame test now uses left_camera as the missing camera
- setup_demo call check loosened to the caller-provided seed/is_test
bits (full kwargs include the YAML-derived blob)
* fix: integrate PR #3315 review feedback
- ci: add Docker Hub login step, add HF_USER_TOKEN guard on eval step
- docker: tie patches to pinned versions with removal guidance, remove
unnecessary HF_TOKEN for public dataset, fix hadolint warnings
- docs: fix paper link to arxiv, add teaser image, fix camera names
(4→3 cameras), fix observation dims (480x640→240x320)
* fix(docs): correct RoboTwin 2.0 paper arxiv link
* fix(docs): use correct RoboTwin 2.0 teaser image URL
* fix(docs): use plain markdown image to fix MDX build
* ci(robotwin): smoke-eval 10 tasks instead of 5
Broader coverage on the RoboTwin 2.0 benchmark CI job: bump the smoke
eval from 5 tasks to 10 (one episode each). Added tasks are all drawn
from ROBOTWIN_TASKS and mirror the shape/complexity of the existing
set (simple single-object or single-fixture manipulations).
Tasks now run: beat_block_hammer, click_bell, handover_block,
open_laptop, stack_blocks_two, click_alarmclock, close_laptop,
close_microwave, open_microwave, place_block.
`parse_eval_metrics.py` reads `overall` for multi-task runs so no
parser change is needed. Bumped the step name and the metrics label
to reflect the 10-task layout.
* fix(ci): swap 4 broken RoboTwin tasks in smoke eval
The smoke eval hit two upstream issues:
- `open_laptop`: bug in OpenMOSS/RoboTwin main — `check_success()` uses
`self.arm_tag`, but that attribute is only set inside `play_once()`
(the scripted-expert path). During eval `take_action()` calls
`check_success()` directly, hitting `AttributeError: 'open_laptop'
object has no attribute 'arm_tag'`.
- `close_laptop`, `close_microwave`, `place_block`: not present in
upstream RoboTwin `envs/` at all — our ROBOTWIN_TASKS tuple drifted
from upstream and these names leaked into CI.
Replace the four broken tasks with upstream-confirmed equivalents
that exist both in ROBOTWIN_TASKS and in RoboTwin's `envs/`:
`adjust_bottle`, `lift_pot`, `stamp_seal`, `turn_switch`.
New 10-task smoke set: beat_block_hammer, click_bell, handover_block,
stack_blocks_two, click_alarmclock, open_microwave, adjust_bottle,
lift_pot, stamp_seal, turn_switch.
* fix(robotwin): sync ROBOTWIN_TASKS + doc with upstream (50 tasks)
The local ROBOTWIN_TASKS tuple drifted from upstream
RoboTwin-Platform/RoboTwin. Users passing names like `close_laptop`,
`close_microwave`, `dump_bin`, `place_block`, `pour_water`,
`fold_cloth`, etc. got past our validator (the names were in the
tuple) but then crashed inside robosuite with a confusing error,
because those tasks don't exist in upstream `envs/`.
- Replace ROBOTWIN_TASKS with a verbatim mirror of upstream's
`envs/` directory: 50 tasks as of main (was 60 with many
stale entries). Added a `gh api`-based one-liner comment so
future bumps are mechanical.
- Update the `60 tasks` claims in robotwin.mdx and
RoboTwinEnvConfig's docstring to `50`.
- Replace the stale example-task table in robotwin.mdx with ten
upstream-confirmed examples, and flag `open_laptop` as
temporarily broken (its `check_success()` uses `self.arm_tag`
which is only set inside `play_once()`; eval-mode callers hit
AttributeError).
- Rebuild the "Full benchmark" command with the actual 50-task
list, omitting `open_laptop`.
* test(robotwin): lower task-count floor from 60 to 50
ROBOTWIN_TASKS was trimmed to 50 tasks (see comment in
`src/lerobot/envs/robotwin.py:48`), but the assertion still
required ≥60, causing CI failures. Align the test with the
current upstream task count.
* fix(envs): preserve AsyncVectorEnv metadata/unwrapped in lazy eval envs
Port of #3416 onto this branch.
* ci: gate Docker Hub login on secret availability
* fix: integrate PR #3315 review feedback
- envs(robotwin): default `observation_height/width` in
`create_robotwin_envs` to `DEFAULT_CAMERA_H/W` (240/320) so they
match the D435 dims baked into `task_config/demo_clean.yml`.
- envs(robotwin): resolve `task_config/demo_clean.yml` via
`CONFIGS_PATH` instead of a cwd-relative path; works regardless
of where `lerobot-eval` is invoked.
- envs(robotwin): replace `print()` calls in `create_robotwin_envs`
with `logger.info(...)` (module-level `logger = logging.getLogger`).
- envs(robotwin): use `_LazyAsyncVectorEnv` for the async path so
async workers start lazily (matches LIBERO / RoboCasa / VLABench).
- envs(robotwin): cast `agent_pos` space + joint-state output to
float32 end-to-end (was mixed float64/float32).
- envs(configs): use the existing `_make_vec_env_cls(use_async,
n_envs)` helper in `RoboTwinEnvConfig.create_envs`; drop the
`get_env_processors` override so RoboTwin uses the identity
processor inherited from `EnvConfig`.
- processor: delete `RoboTwinProcessorStep` — the float32 cast now
happens in the wrapper itself, so the processor is redundant.
- tests: drop the `TestRoboTwinProcessorStep` suite; update the
mock obs fixture to use float32 `joint_action.vector`.
- ci: hoist `ROBOTWIN_POLICY` and `ROBOTWIN_TASKS` to job-level
env vars so the task list and policy aren't duplicated across
eval / extract / parse steps.
- docker: pin RoboTwin + CuRobo upstream clones to commit SHAs
(`RoboTwin@0aeea2d6`, `curobo@ca941586`) for reproducibility.
* fix(one shot load): adding metadata loading when reading from a dataset after writing
* refactor(one shot load): move metadata reload to ensure_readable() on LeRobotDatasetMetadata
Move the metadata reload from DatasetReader.load_and_activate() to a new
public ensure_readable() method on LeRobotDatasetMetadata, called from
LeRobotDataset._ensure_reader(). This places lifecycle management in the
right layer: metadata owns its readiness check, the dataset orchestrates
the write-to-read transition, and the reader stays clean.
Also adds a regression test using delta_timestamps to exercise the
meta.episodes access path in the create -> write -> finalize -> read flow.
Co-authored-by: Steven Palma <imstevenpmwork@users.noreply.github.com>
---------
Co-authored-by: claude[bot] <41898282+claude[bot]@users.noreply.github.com>
Co-authored-by: Steven Palma <imstevenpmwork@users.noreply.github.com>
pepijn