select_message's bf16 cast used next(paligemma.parameters()).dtype,
which lands on a fp32-kept param (norm / embedding) under
to_bfloat16_for_selected_params. Mask stayed fp32 while q/k/v were
bf16 → SDPA still raised "invalid dtype for bias". Read the dtype
from layers[0].self_attn.q_proj.weight instead — q_proj is always
cast with the rest, so its dtype matches what SDPA sees.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
`_prepare_attention_masks_4d` always returns fp32 (the 0.0 / -inf
literals); with bf16 weights, HF PaliGemma's SDPA path raises
"invalid dtype for bias - should match query's dtype" and
select_message returns empty every step. Cast in both attention
sites: `_compute_layer_ki` (training, when both experts run) and
`select_message` (inference, VLM-only branch). Bf16 training +
bf16 inference now run end to end with no dtype mismatch.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Match the production target used in examples/annotations/run_hf_job.py.
Per Scale Labs' dense-captioning ablations, model capacity dominates
prompt-engineering gains; defaulting to the larger model avoids
shipping a worst-tier configuration out of the box.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
THE bug behind the <loc>-salad. PaliGemma's vocab reserves ids
[256000, 257023] for <locDDDD> detection / pointing tokens, but the
stock AutoTokenizer does NOT match them on raw text — it BPE-splits
<loc0162> into SEVEN pieces (<, loc, 0, 1, 6, 2, >). So a VQA target
like "<loc0162><loc0759> green box<eos>" tokenized to 16 pieces, not
5, and training the LM head supervised those generic BPE pieces
instead of one detection-vocab id. The piece logits got pumped up
across ~25% of supervised positions; at inference they dominated
every turn — even subtask prompts produced <loc>-salad followed by
the actual answer.
Register the 1024 <locDDDD> tokens via tokenizer.add_tokens once on
load, in every path the policy uses: PI052TextTokenizerStep (training
encode), _build_text_batch_pi052 (runtime encode), and
select_message's default tokenizer (runtime decode). Verified
empirically with the real PaliGemma tokenizer: VQA target now
tokenizes to 5 ids matching the loc-vocab range (256162, 256759, ...)
with correct offset_mapping.
This unlocks PaliGemma's actual detection prior; <loc>-salad cannot
recur because each <locDDDD> is a single class on the LM head, not a
character sequence the head accidentally learns to extend.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Confirmed empirically on the published dataset: VQA bbox/keypoint
coordinates are Qwen2.5-VL's 0–1000 normalized grounding output, NOT
pixels. Scanning 8207 samples showed x and y both spanning 0..1000
with ~30% of values exceeding the camera's pixel dimensions (which is
impossible if they were pixels).
_vqa_answer_to_loc was dividing by the observation image's H/W, so
e.g. point [742, 158] on a 640x480 wrist cam clamped x to <loc1023>
(the far-right edge) instead of mapping to <loc0760> (~74% across).
Fix: divide by 1000 — the actual Qwen scale. The conversion is now
camera-resolution-independent, so _camera_image_shapes and the
image_shapes plumbing through __call__ / _encode_messages /
_messages_vqa_to_loc are dropped. Tests updated to the new signature
and the 0–1000 round-trip.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
base.fit() rejected the data with "Vocab size 1024 is too small for
the range of tokens 9339": the FAST tokenizer was fit on raw
motor-unit actions, whose DCT-token range vastly exceeds the 1024
codebook.
Two problems, one fix. (1) Raw actions blow up the token range. (2) At
training time ActionTokenizerProcessorStep runs after the QUANTILES
NormalizerProcessorStep, so it encodes normalized actions — fitting on
raw actions mismatches that space. Replicate QUANTILES normalization
(per-dim [q01,q99] -> [-1,1], clipped) before base.fit() so the fit and
the training-time encode see the same distribution and the token range
fits the codebook.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
fit_fast_tokenizer collected action chunks via ds[i]["action"], which
builds a full training item — delta-timestamp expansion, video decode,
image transforms. A single video-decode failure threw, was swallowed
at debug level, and silently starved the fit of every chunk → "FAST
fit collected zero action chunks", falling back to the universal
tokenizer.
Read the ``action`` column straight from the HF dataset instead: it
carries no video, so it is immune to decode errors and far faster.
Also fail fast with a clear message when the dataset has no ``action``
feature or all episodes are shorter than chunk_size.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
_compute_layer_ki called modeling_gemma._gated_residual, but that
adaRMSNorm gated-residual helper is a lerobot helper in pi_gemma, not
part of HF transformers — so enabling knowledge_insulation crashed with
AttributeError on the first training step. Import _gated_residual from
pi_gemma, matching pi05's own layer code.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Spatial VQA answers (bbox / keypoint) were trained as pixel-coordinate
JSON, which fights PaliGemma's detection prior and leaks <loc>-token
salad at inference. Convert them to PaliGemma's native <locNNNN>
vocabulary instead so the LM head reuses that prior.
Training side (text_processor_pi052.py): a target turn whose content
parses as a bbox/keypoint answer is rewritten to <loc> text, using the
camera frame's native (H, W) from the observation and the preceding
image block. Non-spatial answers, subtask/memory targets and SmolVLA2
keep their JSON form — the dataset stays backbone-agnostic.
Runtime side (smolvla2/inference/vqa.py): parse_vqa_answer detects
<loc> answers (2 locs -> keypoint, 4 -> bbox), returning normalized
[0,1] coords with a normalized flag; draw_vqa_overlay denormalizes
against the chosen camera frame's pixel size.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Keep action-only samples trainable by rendering the task as a low-level user message when no recipe branch matches.
Co-authored-by: Cursor <cursoragent@cursor.com>
PI052TextTokenizerStep masked text_labels over the assistant turn's
*content only* — the trailing newline was excluded and no EOS token was
ever a supervised label. So the LM head was never given a stop signal:
at inference select_message decoded to max_new_tokens, producing the
runaway subtask paragraphs and the "}"}"}-style VQA tails.
_format_messages now appends the tokenizer's EOS to each supervised
target turn and extends that turn's span to cover it, so the EOS lands
in text_labels. _shifted_ce then trains "<last content token> -> EOS"
and the model learns to terminate; select_message stops on it.
Inference callers (the runtime's _build_text_batch_pi052) pass no
target_indices / eos_token, so no EOS is baked into the prompt — the
model generates it. Verified end-to-end with the PaliGemma tokenizer:
the supervised span is `<content><eos>` and the trailing newline stays
unsupervised.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
VideoDecoderCache used an unbounded dict keyed on absolute path, with no
eviction in the standard LeRobotDataset path. With shuffled iteration over
datasets that have many distinct mp4 files, every DataLoader worker
accumulated one cached (VideoDecoder, fsspec file handle) pair per distinct
path it had ever touched. Per-entry cost is ~3-5 MB of host RAM plus one
open FD; at ~8 k entries this is roughly 30 GB per worker.
This was hit in the wild during a SmolVLA training run on a 4,195-episode
SO-101 dataset (8,390 mp4s, two cameras per episode). dmesg showed
anon-rss climbing to 34.9 GB on a single pt_data_worker before the OOM
killer fired ~30 min into training; with --num_workers=8 the per-worker
peak halved to 17.9 GB, which is the expected inverse-scaling signature
when the leak is per-decode and the workload is split across workers. The
working workaround on the affected platform was --dataset.video_backend=pyav,
because the pyav path opens/closes per call and never touches this cache.
Switch the backing store to an OrderedDict and evict LRU entries when the
cap is reached, closing the evicted file handle inside the lock so we do
not leak FDs either. Default cap is DEFAULT_DECODER_CACHE_SIZE = 100,
overridable via LEROBOT_VIDEO_DECODER_CACHE_SIZE or by passing max_size=
to the constructor; max_size=None restores the legacy unbounded behaviour
for callers that need it.
Validation on the original failing workload (decode_video_frames_torchcodec
called over real mp4s from the affected SO-101 dataset):
unbounded: 300 files -> +1087 MB host RSS, cache=300, still climbing
cap=50: 500 files -> +266 MB host RSS, cache=50, stable
cap=50: 2000 calls -> +312 MB host RSS, cache=50, stable
cap=100: 1000 calls -> +470 MB host RSS, cache=100, stable
Three independent seeded runs at cap=50 agreed to within 1% (263 / 266 /
265 MB delta), and the 2000-call multi-pass run shows RSS plateaus after
the cap is reached instead of drifting.
Tests in tests/datasets/test_video_decoder_cache.py cover:
default-is-bounded, size cap, LRU ordering, FD close on eviction, FD close
on clear(), cache-hit invariance, max_size=None fallback, and env-var
override. No regressions in test_video_encoding.py, test_streaming.py, or
test_dataset_reader.py (73 prior tests still pass alongside the 8 new ones).
* feat(utility): adding video re-encode utility
* feat(edit): adding a new lerobot-edit-dataset tool to re-encode all the videos of a dataset
* chore(format): formatting code
* chore(review): fix Claude reviews
* test(reencode dataset): adding missing test for reencode dataset
* Add extensive language support
* Address review: split persistent/event schemas, drop event timestamps
- recipe.py: derive _VALID_ROLES/_VALID_STREAMS from MessageRole/MessageStream Literals
- dataset_metadata.py: keep CODEBASE_VERSION at v3.0
- language.py: remove RESERVED_STYLES; split arrow/feature schemas into
persistent (with timestamp) and event (without timestamp); add docstrings
- language_render.py: events use frame-row timestamp implicitly; no
per-event timestamp filtering or sorting
- converters.py: drop unused subtask_key passthrough
- add docstrings to new public APIs (recipe, render_messages_processor, collate)
- update tests for split schemas; revert uv.lock
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Add docstrings to all new helpers; revert uv.lock
Covers private helpers in recipe.py, language.py, language_render.py,
and render_messages_processor.py. Also reverts uv.lock to main (it was
re-generated by `uv run` during local checks).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(language): add motion (persistent) and trace (event-only) styles
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(language): per-camera tagging on view-dependent styles
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>
* fix(language): drop motion from VIEW_DEPENDENT_STYLES
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>
* feat(language): task_aug style + automatic ${task} rephrasing rotation
Adds task-prompt diversity (Xiao 2022 / CAST) without touching
``meta/tasks.parquet`` or forcing recipes to opt in. The plan reserved
``task_aug`` as a future style; this lands it now.
- ``language.py``: add ``task_aug`` to ``CORE_STYLES`` and
``PERSISTENT_STYLES``. ``column_for_style("task_aug")`` returns
``language_persistent`` so PR 2 writers route it correctly.
- ``language_render.py``: ``_resolve_task`` now consults the persistent
slice for rows of ``style="task_aug", role="user"``. When any exist
it picks one deterministically by ``sample_idx`` (blake2b-keyed, not
Python's randomized hash) so an epoch sees every rephrasing of every
episode while the same sample still resolves identically across
reruns. Falls back to the canonical ``meta/tasks.parquet`` task when
no rephrasings are present, so existing datasets and unannotated runs
keep their behaviour. Explicit ``task=`` overrides still win.
- Tests: rephrasing coverage across samples, determinism on repeat
``sample_idx``, fallback when persistent has no ``task_aug`` rows,
and explicit override priority.
Recipes get this for free: any ``${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>
* feat(language): tool catalog in meta/info.json + LeRobotDatasetMetadata.tools
Stores OpenAI-style function schemas at ``meta/info.json["tools"]`` so
datasets can declare which tools are available (today: just ``say``;
tomorrow: per-dataset extensions). The ``DEFAULT_TOOLS`` constant
fills in for unannotated datasets so chat-template consumers don't
have to special-case anything.
Three pieces:
- ``language.py``: ``SAY_TOOL_SCHEMA`` and ``DEFAULT_TOOLS``
constants. Single source of truth — PR 2's writer and PR 3's
runtime tool registry will both import from here instead of
duplicating the dict.
- ``dataset_metadata.py``: ``LeRobotDatasetMetadata.tools`` property
reads ``info.json["tools"]`` and falls back to ``DEFAULT_TOOLS``.
Returns deep-copied dicts so callers can mutate the result safely.
- ``docs/source/tools.mdx``: spec page covering the catalog, per-row
invocations, and the three-step "how to add a new tool" workflow
(declare schema, implement, register). Linked from the docs
toctree under the Datasets section.
This lays the groundwork for PR 2's pipeline writing the catalog out
during annotation, and PR 3's ``src/lerobot/tools/`` package shipping
runnable implementations (one file per tool — first up:
``say.py`` wrapping Kyutai's pocket-tts).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Apply ruff and prettier formatting after merge
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* refactor(language): unify resolver dispatch and prune redundant test scaffolding
* 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>
* fix(language): always raise on ambiguous resolver matches
`_select_one` previously skipped its ambiguity check whenever any of
`role`/`tool_name`/`camera` was set, on the assumption that the caller
had already pinned down a unique row. That left a real ambiguity hole
for VQA: with two cameras emitting `(vqa, assistant)` at the same
frame, `emitted_at(..., role="assistant")` silently picked the first
sorted row instead of telling the recipe to add `camera=...`. The
existing `test_emitted_at_raises_on_ambiguous_per_camera_vqa` test
already encoded the desired behavior.
Tighten the check: any time `len(rows) > 1` we now raise with the
selectors echoed back, so users see exactly which fields they passed
and that more is needed to disambiguate.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* chore: fix CI — collapse short ValueError to one line, refresh uv.lock
* `ruff format` on CI (newer version) wants the short `camera=None`
ValueError on a single line.
* `uv.lock` was stale relative to `pyproject.toml`'s `datasets>=4.7.0`
pin (and picked up upstream `s390x` marker fixes for cuda packages).
CI runs `uv sync --locked` which rejected the divergence.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(language): keep base install green — drop processor re-export, gate dataset-extra tests
`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>
* fix(language): address review — tools accessor, motion docs, conditional collate
* **`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>
* review: dedupe regex, centralize column names, harden collate, more tests
* **#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>
* review: emitted_at uses 0.1s tolerance; MessageTurn requires stream at construction
* **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>
* docs(tools): drop follow-up-PR references
Reword the two callouts in `tools.mdx` to describe the runtime layer
in present tense ("not part of the catalog layer shipped today",
"those modules don't yet exist in the tree") instead of pointing at a
specific follow-up PR. Keeps the doc honest about what works now
without coupling it to a particular release order.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* review: address CarolinePascal feedback
- language timestamps: float64 -> float32 to match LeRobotDataset frame
timestamps (Arrow struct + HF feature)
- dataset_metadata: hoist `.language` imports to module top — language.py
has no lerobot imports, so there is no circular-import risk
- dataset_metadata: add a `meta.tools` setter that persists the catalog to
info.json and reloads `meta.info`
- feature_utils: validate the `language` dtype instead of returning "" —
warn (non-fatal) when a non-empty value is written at record time
- centralize the scalar-unwrap helper as `lerobot.utils.utils.unwrap_scalar`,
shared by render_messages_processor and language_render
- docs: move `## Layer 2 — recipe anatomy` ahead of the resolver sections,
which describe recipe bindings rather than dataset layout
- language_render: note in EMITTED_AT_TOLERANCE_S that persistent rows change
on a human-action timescale, not the camera frame rate
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The runtime's text path was hard-wired to SmolVLA2: _build_text_batch
read policy.config.vlm_model_name (which PI052Config doesn't have) and
built a SmolVLM2 chat-template prompt. PI052/PaliGemma is not
chat-pretrained and trains on a flat `User: ... \nAssistant: ...`
prompt, so the runtime crashed or fed an out-of-distribution prefix.
- _build_text_batch now dispatches on policy.config.type: smolvla2 ->
chat template (renamed _build_text_batch_chat); pi052 -> flat
role-prefixed text via PI052TextTokenizerStep's own _format_messages /
_strip_blocks / _flatten_say_tool_calls, so the inference prefix
matches PI052 training exactly.
- Add a lerobot-pi052-runtime entry point (alias of the same main; the
policy type is read from the checkpoint) so the command name isn't
misleading. argparse prog now defaults to the invoked command name.
PI052's select_message / predict_action_chunk already work with the
runtime; this was the one SmolVLA2-only coupling.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The first-person memory narrative, task-rephrasing and initial-speech
prompt tweaks belong in the annotation pipeline itself. Applied to
feat/language-annotation-pipeline (#3471); reverting them here to the
merge-base so they drop out of this PR's diff. general_vqa.py keeps its
docstring fix since it references a recipe this PR introduces.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- module_1_memory: rewrite as an explicit first-person, past-tense
narrative ("I picked up...", "I opened...") matching the MEM
(Torne 2026) running-memory style, instead of "one or two short
sentences" with no person/tense guidance.
- module_1_task_rephrasings: bias rephrasings toward short imperative.
- module_2_initial_speech: prefer very short robot acknowledgements.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The deterministic-plan rewrite, single-frame VQA (K 3->1), dataset
version tagging, telegraphic-subtask prompt and shorter interjection
prompt belong in the annotation pipeline itself, not in the SmolVLA
training PR. They have been applied to feat/language-annotation-
pipeline (#3471). Reverting these six files here to the merge-base so
they drop out of this PR's diff; #3491 will inherit the canonical
versions when it next rebases on its base.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.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>
Rewrite module_1_subtasks prompt to produce short imperative commands
("pick up the orange") instead of third-person narration ("the robot
arm moves to the orange"). Drops the verbose "how, not what" rule and
adds a good/bad few-shot table.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Convert PI052 prefix-only attention masks before calling PaliGemma so text-only batches and generation use the same mask shape as fused training.
Co-authored-by: Cursor <cursoragent@cursor.com>
Route full PI05/PI052 fine-tuning through PyTorch's fused AdamW path to avoid the single-tensor Adam denominator allocation near GPU memory limits.
Co-authored-by: Cursor <cursoragent@cursor.com>
Avoid the multi-tensor AdamW temporary that can OOM full PI05/PI052 fine-tuning near GPU memory limits.
Co-authored-by: Cursor <cursoragent@cursor.com>
Select only supervised text and FAST action-code positions before cross-entropy to avoid full-vocabulary loss tensors over padded sequences.
Co-authored-by: Cursor <cursoragent@cursor.com>
Use PI05Policy helpers for action padding and image preprocessing in PI052 fused losses instead of looking them up on the inner PI05Pytorch module.
Co-authored-by: Cursor <cursoragent@cursor.com>
* feat(robots): natively integrate Seeed Studio reBot B601-DM arm
Add first-class LeRobot support for the Seeed Studio reBot arm, replacing
the out-of-tree `lerobot-robot-seeed-b601` / `lerobot-teleoperator-rebot-arm-102`
plugin packages.
New devices:
- robot `rebot_b601_follower` — single-arm B601-DM follower (6-DOF + gripper,
Damiao CAN motors via `motorbridge`)
- robot `bi_rebot_b601_follower` — bimanual follower composing two single arms
- teleoperator `rebot_102_leader` — single-arm StarArm102 / reBot Arm 102 leader
(FashionStar UART servos via `motorbridge-smart-servo`)
- teleoperator `bi_rebot_102_leader` — bimanual leader composing two single arms
The bimanual variants reuse the single-arm classes and namespace each arm's
observation/action keys with `left_` / `right_` prefixes, so a bimanual
StarArm102 leader can teleoperate a bimanual reBot B601 follower.
Optional SDK imports are guarded; a `rebot` extra installs `motorbridge` and
`motorbridge-smart-servo`.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs: add reBot B601-DM calibration & dual-arm teleoperation guide
Add docs/source/rebot_b601.mdx covering single-arm and bimanual
calibration and teleoperation for the reBot B601-DM follower and
reBot Arm 102 leader, with zero-position reference images from the
Seeed Studio wiki. Register the page in the docs toctree.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs: fix reBot B601 MDX build (move JSON example out of <Tip>)
The doc-builder parses `{...}` inside MDX component children as a
Svelte expression, so the joint_directions JSON example broke the
build. Move it into a top-level fenced code block.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs: apply prettier formatting to reBot B601 page
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs: remove duplicate colocated reBot B601 page
docs/source/rebot_b601.mdx is the canonical, toctree-registered page;
the colocated rebot_b601.md was a redundant thinner copy.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs: clarify 6-DOF leader fallback comment in reBot B601 follower
Explain that holding wrist_yaw at zero is what lets a 6-DOF leader
(e.g. so100_leader / so101_leader) teleoperate the 7-DOF follower.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* refactor: address Caroline's PR review on reBot B601 integration
- leader: remove _validate_config (no other lerobot device validates its
config; a key mismatch now surfaces as a plain KeyError)
- leader: simplify _round_to_valid_range to direct modular arithmetic
instead of a bidirectional search loop
- leader: inline the single-use _clamp helper
- follower & leader: write MotorCalibration range_min/range_max from the
configured joint_limits / joint_ranges instead of a fixed [-90, 90]
- docs: add a "Find the USB ports" section (lerobot-find-port) and move
the brltty/permissions tip there; link the OpenArm page for SocketCAN
adapter configuration
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Tokenize batched recipe outputs in PI052 so training batches with nested message lists do not crash before model forward.
Co-authored-by: Cursor <cursoragent@cursor.com>
Mask the FAST auxiliary loss to discrete action-code tokens so wrapper formatting tokens do not affect action co-training.
Co-authored-by: Cursor <cursoragent@cursor.com>
Module 3 anchored each VQA emission tick to K=3 consecutive frames
(~0.1s at 30fps). The VLM grounds the answer — bbox/keypoint
coordinates especially — against the first frame's image, so copying it
onto frames 2-3 smears a stale label over a moving scene.
Default K=1: a VQA pair lands on exactly its emission frame, no
temporal smear. VQA frames get sparser; the WeightedEpisodeAwareSampler
(vqa_target_fraction) is the knob to compensate.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The decoder chain tried torchcodec first, then ffmpeg. torchcodec is
not thread-safe: under the executor's 16-wide concurrent decode in the
interjections phase it SIGSEGVs (exit 139) before the ffmpeg fallback
is ever reached — uncatchable, so it kills the whole job.
Default the auto chain to ffmpeg only. Per-frame ffmpeg decode runs in
an isolated child process: crash-safe and concurrency-safe (the plan
phase already proved 16 parallel ffmpeg subprocesses are fine).
torchcodec / pyav remain available via an explicit video_backend.
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>
- hirobot.yaml -> subtasks_vqa.yaml
- hirobot_memory.yaml -> subtask_mem_vqa_speech.yaml
- pi05_hirobot.yaml -> deleted (stale: uses plan, top-camera names;
superseded by the two recipes above)
- smolvla2_hirobot.yaml -> deleted (was untracked stale junk)
Updated the smolvla2 / pi052 `recipe_path` config defaults, all
docstring / comment references, the annotation-pipeline + recipe docs,
and the three tests that loaded pi05_hirobot.yaml (repointed to the
renamed recipes; the low-level-branch and pipeline-render assertions
now accept a flow-only `low_level` stream as valid supervision, since
the new recipes' low_level_execution has no text-CE target).
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>
pi052 had the same text-CE collapse bug smolvla2 had — PaliGemma's
embed_prefix flags the language block att=0, so make_att_2d_masks makes
it fully bidirectional and the text cross-entropy degenerates into a
copy task. Ported the three model-specific fixes:
- _mark_target_span_causal: set att=1 on supervised target language
positions so the text-CE is genuine causal next-token prediction.
Applied in both _compute_all_losses_fused and _compute_text_and_fast_loss.
- flow_loss_weight 10.0 -> 5.0: the paper's a=10 swamps the LM head once
the flow-only low_level recipe fires often (matches SmolVLA2Config).
- _flatten_say_tool_calls in the text tokenizer: serialize `say` tool
calls into a <say>...</say> marker so the spoken reply is tokenized
and supervised (PaliGemma's flat prompt has no structured calls, so
they were dropped entirely).
select_message needed no change: pi052's prefix is [images, language]
with no trailing state token, so it already decodes from the last
language token.
Regression tests mirror the smolvla2 attention-masking + tool-call suite.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
VQA annotations are sparse, so VQA was badly underrepresented in training:
its effective share was weight x density, and blend draws that picked an
ask_vqa* sub-recipe for a non-VQA frame were wasted entirely.
Two pieces:
1. Recipe-side consumption (language_render.py): render_sample now routes
any frame that carries a VQA annotation to a matching ask_vqa* sub-recipe,
regardless of the weighted blend draw. No VQA annotation is wasted and no
draw lands on a non-renderable VQA recipe — VQA's recipe-side share now
equals the VQA-annotation density.
2. Dataset-side oversampling (WeightedEpisodeAwareSampler + vqa_target_fraction):
a new weighted, episode-aware sampler draws frames with replacement by
per-frame weight. When TrainPipelineConfig.vqa_target_fraction is set, the
train script scans language_events, weights VQA frames so they make up
~that fraction of the training stream, and uses the weighted sampler. This
is what actually lets VQA exceed its natural density. Default None keeps
uniform episode-aware sampling unchanged.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
VideoFrameProvider decoded keyframes via torchcodec only. Some containers
(e.g. vllm-openai) ship a torchcodec that cannot push packets to the
decoder ("Operation not permitted"), silently degrading interjection/vqa
prompts to no visual context.
_decode now retries with pyav when the default backend raises, and a new
`video_backend` config field lets callers pin the backend explicitly.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds an optional `dest_repo_id` to AnnotationPipelineConfig. When set,
`push_to_hub` uploads the annotated dataset there instead of overwriting
the source `repo_id`, restoring separate source/destination repos.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Enhance documentation with Lance format details
Added information about Lance format and `lerobot-lancedb` package for multimodal AI datasets.
Signed-off-by: Quentin Lhoest <42851186+lhoestq@users.noreply.github.com>
The VLM already sees every camera, so the operator never needs to name
one to ask a question. Move the camera prompt to after generation and
only fire it when the answer actually carries a bounding box / point
(whose pixel coordinates are camera-specific and need a target frame).
Non-spatial answers (count / attribute / spatial / plain text) now skip
the prompt entirely.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
handle_vqa_query filtered the observation down to the single chosen
camera before calling the VLM. But training feeds every camera: the
ask_vqa_* recipes' image blocks are stripped before tokenization and
the frames reach the model via OBS_IMAGES_*, where embed_prefix
consumes all config.image_features regardless of the per-camera recipe
tag. Filtering to one camera changed the image-token count in the
prefix (the dropped camera zero-padded with mask=0) — a prefix shape
the model never saw at training.
Now the full observation is passed to select_message; the chosen
camera is used only to pick which frame the bbox/point overlay is
drawn on.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Command arguments never needed quotes (`_strip_quotes` only strips a
matching pair if present) — `/question point to the yellow cube` works.
The hints wrongly implied `""` were required; all hints/help now show
`/action <task>` / `/question <text>`.
Also adds a reference line to the state panel showing the two
overlay-producing VQA prompt shapes:
/question point to the yellow cube -> point overlay
/question detect the blue cube -> bounding-box overlay
plus the same examples in /help.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replace the startup mode prompt + task picker with a single
command-driven prompt. The runtime now comes up immediately at the
command line in `paused` mode (robot idle) and the operator drives it:
/action "task" run the robot on a task (bare = resume, number = timed burst)
/pause stop the action loop — robot holds position
/question "..." pause and answer one VQA question (camera prompt + overlay)
/help / stop
- Removed _select_mode_interactively / _select_task_interactively /
_dataset_task_strings (the interactive pickers).
- mode value renamed "question" -> "paused"; --mode choices are now
action|paused (default paused).
- /question takes the question inline and runs it via _handle_slash_command
(pauses first, so the policy isn't used concurrently).
- The ENTER-to-start gate only fires when starting in action mode.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Lets the operator skip the interactive startup entirely and go straight
to the command line:
- New --mode {action,question} arg; when given, the startup mode prompt
is skipped.
- When --task is passed explicitly on the CLI, the startup task picker
is skipped (the dataset-bootstrap task still shows the picker so you
can override it).
Also adds a timed action burst: /action <seconds> runs the robot for N
seconds, then the autonomous loop auto-reverts to question mode and
clears the action queue. Plain /action stays unlimited.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
New recipe alongside hirobot.yaml (kept as the lean baseline). Superset
that adds two text-supervised sub-recipes:
- memory_update: compress progress into a memory note.
- user_interjection_response: reply to a user interjection with a `say`
tool call only (no plan/subtask text). The SmolVLA2 chat tokenizer
flattens the call to a `<say>...</say>` marker the runtime parses back.
Plan is intentionally omitted; memory is the only persistent high-level
state. Weights: low_level 0.40, subtask 0.25, memory 0.10, interjection
0.10, vqa 0.075 x2.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add a mode prompt at startup, shown before the task picker, so the
operator chooses action (run the robot) vs question (VQA only) up front
instead of having to discover /vlm mid-run.
Also rename the VQA mode from "vlm" to the clearer "question":
- state["mode"] value is now "action" | "question"
- the command is /question (/vlm and /vqa kept as aliases)
- panels, hints and help text updated to match
handle_vqa_query now reports via both push_log and direct stdout, so
VQA answers / overlay paths are visible in autonomous question mode
where the panel redraw is suspended.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The autonomous panel redraw cleared the screen every 0.5s, so the "> "
prompt and the one-shot command hint vanished — the operator could not
see what to type or what they were typing, making /vlm unreachable.
- Suspend the timer redraw entirely while in /vlm mode (the action loop
is paused, nothing changes in the background) so the VQA question and
camera prompt stay on a stable screen.
- Re-print the "> " prompt after each redraw so it is always visible.
- Show an always-on command hint in the panel (/vlm, /help, /action)
instead of relying on the startup line that scrolls away.
- Redraw immediately after a slash command so the mode flip is visible.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Pepijn