* feat(train): FSDP checkpoint saving
* adding docs for FSDP
* adding a test for the fsdp checkpoint path
* cleanup
* fixing final upload to hub
* refactored initial implementation to use torch fsdp api and adding new tests
* fix(datasets): enforce one parquet row group per episode in v3 data writes
LeRobot v3 data shards must hold exactly one row group per episode so a
reader can fetch episode i with pq.ParquetFile(path).read_row_group(i)
(a byte-range read) instead of loading the whole shard. The recording
writer already does this (one write_table per episode); the aggregate
and lerobot-annotate re-write paths instead concatenated many episodes
and wrote them in one shot, collapsing the file to a single row group.
- io_utils: add write_table_one_row_group_per_episode (one ParquetWriter,
one write_table per episode — same pattern as the recording writer);
to_parquet_with_hf_images embeds images then writes per-episode row
groups; to_parquet_one_row_group_per_episode wraps it for plain frames
- aggregate: route non-image data writes through the per-episode writer;
leave the episodes-metadata parquet untouched (already one row/episode)
- annotate: rewrite shards via the per-episode writer instead of a single
bulk pq.write_table
- tests: invariant coverage through the aggregate (image + video) and
annotate paths
No change to on-disk schema, paths, naming, rollover thresholds, or
compression. Readers stay backward-compatible (old collapsed files load).
* Update src/lerobot/datasets/io_utils.py
Co-authored-by: Caroline Pascal <caroline8.pascal@gmail.com>
Signed-off-by: Pepijn <138571049+pkooij@users.noreply.github.com>
* Update src/lerobot/datasets/io_utils.py
Co-authored-by: Caroline Pascal <caroline8.pascal@gmail.com>
Signed-off-by: Pepijn <138571049+pkooij@users.noreply.github.com>
* fix(datasets): correct indentation and add strict= in row-group helper
The web-edited numpy version of write_table_one_row_group_per_episode had an
over-indented line (IndentationError, breaking pre-commit + test collection)
and a zip() without strict=. Fix both; behaviour unchanged.
---------
Signed-off-by: Pepijn <138571049+pkooij@users.noreply.github.com>
Co-authored-by: Caroline Pascal <caroline8.pascal@gmail.com>
* fix(images/videos): fixing aggregate_pipeline_dataset_features to avoid unwanted images features deletion when videos are not used
* fix(docstrings): improving docstrings
Signed-off-by: Caroline Pascal <caroline8.pascal@gmail.com>
---------
Signed-off-by: Caroline Pascal <caroline8.pascal@gmail.com>
* chore(robots): homogenize bi setups
* feat(robots): split openarm mini into single and bi
* refactor(robots): mixin for bi classes
* docs: update docs
* feat(edit-dataset): add `concatenate_videos` opt-out to merge
When merging datasets, source mp4s are concatenated into shards capped at
`video_files_size_in_mb` (default 200 MB). This is great for dataloader
throughput but destroys per-episode (or per-source) video boundaries,
which is undesirable when you want to inspect, ship, or reuse the
individual mp4s.
Add a `concatenate_videos: bool = True` knob plumbed through
`MergeConfig` → `merge_datasets` → `aggregate_datasets` → `aggregate_videos`.
When False, each source mp4 is copied 1:1 to its own destination mp4 with
no re-muxing, so the merge preserves source video boundaries.
Usage:
lerobot-edit-dataset \
--new_repo_id user/merged \
--operation.type=merge \
--operation.repo_ids "['user/a', 'user/b']" \
--operation.concatenate_videos=false
Defaults are unchanged; the dataloader path is unaffected because the
`episodes.parquet` `from_timestamp`/`to_timestamp` index keeps working
regardless of whether each mp4 holds one or many episodes.
* feat(edit-dataset): extend concatenate opt-out to data files
Following review, add a concatenate_data flag mirroring concatenate_videos,
threaded through MergeConfig, merge_datasets, aggregate_datasets, aggregate_data
and append_or_create_parquet_file. Metadata index files still always concatenate.
Also trim the verbose docstrings and comments since the names are
self-explanatory, and extend the existing merge test to cover data files.
Steerable annotation pipeline (lerobot-annotate) that populates the language_persistent and language_events columns introduced in PR 1 (#3467) directly into data/chunk-*/file-*.parquet.
This is PR 2 of the three-PR plan:
PR 1 (Add extensive language support #3467): schema + DSL + rendering, base of this PR
PR 2 (this PR): annotation pipeline writing into PR 1's columns
PR 3: model with language prediction and runtime
A VLM (Qwen-VL family, served on vLLM) watches each episode's video and emits grounded language annotations: subtasks, plans, memory, task rephrasings, interjections + speech, and per-camera VQA. The pipeline is built for production annotation at scale — single-camera grounding, embedded-frame inputs, a describe-then-segment grounding flow, and a deterministic full-episode coverage guarantee — informed by Scale's dense-captioning findings (representation > sampling, rules > reasoning, model capacity is the biggest lever, two-pass systems compound errors)
* fix(datasets): expose a generator on EpisodeAwareSampler for distributed shuffle sync
In distributed training, accelerate can only synchronize the shuffle
permutation across ranks when the sampler exposes a generator attribute.
EpisodeAwareSampler shuffled via the global torch RNG, so disjoint batch
shards relied on every rank's global CPU RNG staying in lockstep forever;
any rank-asymmetric RNG consumption (e.g. eval rollouts on the main
process only) silently desynced the permutations and ranks trained on
overlapping/missing samples.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(train): seed sampler generator and gate dataset download per node
- Pass a generator seeded with cfg.seed to EpisodeAwareSampler so
accelerator.prepare registers it as the synchronized RNG and the
shuffle order is reproducible.
- Gate the initial make_dataset call on is_local_main_process instead of
is_main_process: the global main process only exists on node 0, so on
every other node all local ranks were downloading the dataset and
building the Arrow cache concurrently.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(datasets): add DeterministicEpisodeAwareSampler with O(1) memory and sample-exact resume
Add a sampler that never materializes frame indices: it stores only
per-episode boundaries (numpy, a few bytes per episode) and maps logical
positions to frame indices on the fly with searchsorted. Shuffling uses a
seeded Feistel permutation over [0, num_frames) (cycle-walking to the
exact domain), so the data order is a pure function of (seed, epoch):
- no RNG state to synchronize across distributed ranks,
- constant memory and zero epoch-boundary cost at any dataset size,
- O(1) seek to any position, enabling sample-exact resume.
Opt in with --deterministic_sampler=true. On resume, lerobot-train maps
the checkpointed step back to (epoch, start_index) via
compute_sampler_state and continues at the exact sample where the run
left off (up to accelerate's even_batches padding at epoch boundaries).
The shuffle is pseudo-random rather than a true uniform permutation, the
standard trade-off in large-scale training loaders.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* refactor(datasets): fold deterministic mode into EpisodeAwareSampler
Instead of a parallel DeterministicEpisodeAwareSampler class, extend the
existing EpisodeAwareSampler with a deterministic=True mode (seeded
Feistel permutation, epoch auto-advance, state_dict/load_state_dict).
The default mode is behavior-identical: same torch.randperm consumption
and the same generator contract accelerate synchronizes; the O(N) Python
index list is replaced by O(num_episodes) boundary arrays in both modes,
with `indices` kept as a back-compat property. Passing a generator
together with deterministic=True is rejected, and the state/seek methods
raise outside deterministic mode.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(train): enable deterministic_sampler by default
Deterministic data order (sample-exact resume, no cross-rank RNG sync,
O(1) sampler memory) is now the default for map-style training; set
deterministic_sampler=false to restore the legacy RNG-based shuffle.
Streaming datasets ignore the flag (the sampler path only applies to
map-style datasets), replacing the previous hard validation error so
streaming configs keep working with the new default.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(datasets): default EpisodeAwareSampler to deterministic mode and trim comments
deterministic=True is now the class default as well as the training
default; the legacy RNG path requires an explicit deterministic=False
(the train script's non-deterministic branch passes it). Docstrings and
inline comments slimmed down across the changed files.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test(sampler): drain resumed trillion-frame sampler via iter() to avoid list() prealloc
list(sampler) calls PyObject_LengthHint -> __len__ (the full 10**12 epoch length) and
preallocates that many slots before iterating, OOMing even though the resumed epoch only
yields 3 frames. Collect through the iterator (no length hint) so the test exercises the
real O(1) seek/drain instead of CPython's list growth heuristic.
* fix(datasets): guard Feistel cycle-walking loop against non-convergence
Replace the unbounded while True in EpisodeAwareSampler._permute with a
bounded for loop capped at _MAX_CYCLE_WALK_STEPS (100) and raise
RuntimeError if the cycle-walk fails to land in [0, num_frames). The
loop is expected to converge in <4 steps on the chosen power-of-two
domain, so the bound is a safety net that should never trip in practice
but prevents a pathological infinite loop.
https://claude.ai/code/session_01HQ15tFrBsHYScjGWosEv22
* fix(datasets): make deterministic-sampler resume robust to world-size changes
compute_sampler_state mapped a checkpointed step back to (epoch, start_index)
using the *current* num_processes, but the number of sampler positions a step
consumes scales with the world size that produced it. Resuming on a different
GPU count therefore landed on the wrong epoch/offset, silently re-seeing or
skipping data.
Record num_processes in training_step.json at checkpoint time and feed the
checkpoint's value into compute_sampler_state on resume, so the data order
resumes at the right position regardless of the new world size. Warn when the
world size changed (the global offset is correct, but per-rank sample-exactness
needs the same topology). Old checkpoints without the field fall back to the
current world size.
Also document compute_sampler_state's assumptions explicitly: num_processes /
batch_size must match the checkpointing run, and accelerate's even_batches=True
padding is mirrored by the ceil(... / num_processes) term.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
* style: apply ruff-format to lerobot_train.py
Collapse the compute_sampler_state(...) call onto one line so the
ruff-format pre-commit hook passes (fixes the failing CI check).
Co-authored-by: Cursor <cursoragent@cursor.com>
* refactor(datasets): use seeded torch.randperm instead of Feistel in EpisodeAwareSampler
Drop the Feistel permutation (and its SplitMix64 hash / cycle-walking) in favor of a
torch.randperm seeded from (seed, epoch). The deterministic mode keeps its key properties
- data order is a pure function of (seed, epoch), so it reproduces on every rank with no
global-RNG synchronization, and
- state_dict / load_state_dict still resume sample-exactly, now by regenerating the epoch's
permutation and slicing from the saved offset.
Construction stays O(num_episodes) (only episode boundaries are stored, never a per-frame
index list). The trade-off vs Feistel: the per-epoch shuffle is again O(num_frames) memory
(the randperm tensor) and no longer O(1)-seekable, in exchange for ~30 fewer LOC and a truly
uniform shuffle. Tests updated: the trillion-frame O(1) test is replaced with a
boundary-storage check and a scale resume-exactness test.
Co-authored-by: Cursor <cursoragent@cursor.com>
* refactor(datasets): make EpisodeAwareSampler always deterministic
With Feistel gone, deterministic and legacy modes were both just torch.randperm and the
deterministic path strictly dominated (reproducible across ranks via the (seed, epoch) seed,
no accelerate generator sync, resumable). Collapse to a single path and drop the redundant
flag:
- remove the `deterministic` and `generator` constructor args, `_iter_default`, and
`_require_deterministic`; `set_epoch` / `state_dict` / `load_state_dict` are now unconditional
- remove the `deterministic_sampler` train config field and the legacy generator branch in
lerobot_train.py (non-streaming map datasets always use the sampler)
- drop the now-obsolete generator/legacy tests
Note: removes the `generator` kwarg from EpisodeAwareSampler (back-compat break vs main); the
order is now a pure function of (seed, epoch), so no cross-rank RNG sync is needed.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(datasets): address sampler review (batch_size resume guard + docs)
- Record batch_size in training_step.json alongside num_processes and feed
the checkpoint's value into compute_sampler_state on resume; warn when it
differs (per-rank sample-exactness needs the same batch size).
- Document the set_epoch vs __iter__ auto-advance coupling on EpisodeAwareSampler
(callers should rely on exactly one mechanism per run).
- Note the broadened (reproducibility-breaking) sampler guard and the no-generator
distributed sharding correctness in lerobot_train.py.
- Add load_training_batch_size + parallel tests.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(train): download dataset once on the global main process
Gate the training dataset download on the global is_main_process (download once to the
shared dataset root, barrier, then every other rank reads the already-populated copy)
instead of per-node is_local_main_process. LeRobotDataset skips its snapshot_download
when try_load() succeeds, so no rank re-downloads. Assumes the dataset root / HF cache is
on storage shared across nodes.
Co-authored-by: Cursor <cursoragent@cursor.com>
* chore(datasets): trim sampler comment and drop duplicate tests
Remove the verbose dataloader-guard comment and the two EpisodeAwareSampler tests
that duplicated existing validation/warning coverage (no coverage loss).
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(datasets): expose a generator on EpisodeAwareSampler for distributed shuffle sync
In distributed training, accelerate can only synchronize the shuffle
permutation across ranks when the sampler exposes a generator attribute.
EpisodeAwareSampler shuffled via the global torch RNG, so disjoint batch
shards relied on every rank's global CPU RNG staying in lockstep forever;
any rank-asymmetric RNG consumption (e.g. eval rollouts on the main
process only) silently desynced the permutations and ranks trained on
overlapping/missing samples.
* fix(train): seed sampler generator and gate dataset download per node
- Pass a generator seeded with cfg.seed to EpisodeAwareSampler so
accelerator.prepare registers it as the synchronized RNG and the
shuffle order is reproducible.
- Gate the initial make_dataset call on is_local_main_process instead of
is_main_process: the global main process only exists on node 0, so on
every other node all local ranks were downloading the dataset and
building the Arrow cache concurrently.
* feat(processor): add in-memory pipeline serialization
Expose processor pipeline config and tensor state without requiring temporary files, so processors can be transported, compared, or hashed directly in memory.
* feat(processor): enhance DataProcessorPipeline with registry support
- Added a new RegisteredLazyTensorStateStep for registry-based serialization tests.
- Improved state filename handling in _get_state_filename method.
- Refactored validation logic in _validate_loaded_config to simplify parameter types.
- Updated tests to verify registry step functionality and ensure correct state loading.
* refactor(processor): update state handling in DataProcessorPipeline
- Introduced a new static method _get_state_key to derive in-memory state keys from serialized filenames.
- Updated state_dict and load_state_dict methods to use suffixless state keys instead of filenames.
- Adjusted related tests to reflect changes in state key handling, ensuring consistency in state management
* fix(processor): update loaded_config argument description in DataProcessorPipeline
- Clarified the documentation for the loaded_config parameter to indicate that it may be a non-dictionary value, enhancing understanding for future developers.
Eliminate the standalone pi052/pi05_backbone.py by distributing its contents:
- Generic dual-expert transformer machinery -> lerobot/policies/pi_gemma.py
(sdpa_attention_forward, compute_layer_complete, PaliGemmaWithExpertModel,
get_gemma_config; the openpi width/depth config is renamed GemmaConfig ->
GemmaVariantConfig to avoid clashing with transformers' GemmaConfig). These
sit next to the existing PiGemma layer code they already depend on.
- pi052-specific model + helpers -> pi052/modeling_pi052.py (PI05Pytorch,
ActionSelectKwargs, make_att_2d_masks, pad_vector, resize_with_pad_torch,
create_sinusoidal_pos_embedding, sample_beta, get_safe_dtype).
DEFAULT_IMAGE_SIZE is duplicated as a plain constant in pi_gemma to avoid a
pi_gemma -> pi05 import cycle. Additive to pi_gemma; pi0/pi05 unaffected.
Verified bit-exact on pepijn223/pi052_robocasa_full (embed/predict/forward
identical) and all 34 pi052 tests pass.
Co-authored-by: Cursor <cursoragent@cursor.com>
_fast_ce/_shifted_ce were renamed to _fast_lin_ce/_shifted_lin_ce and changed
from logits-based to Liger fused-linear-CE (hidden @ lm_head_weightᵀ). Update
the tests via thin adapters that pass an identity lm_head_weight (so the
computed logits equal the provided ones), run on CUDA (Liger is GPU-only) and
skip otherwise, and loosen the allclose tolerance to absorb GPU-vs-CPU float
noise on the tiny losses.
Co-authored-by: Cursor <cursoragent@cursor.com>
The smolvla branch had modified the shared pi0/pi05 modeling + pi05 config to
support pi052 (SDPA attention, layernorm/lm_head handling, optimizer
foreach/fused/lm_head_lr_scale, embedding scaling). Decouple pi052 instead:
- Vendor the PI0.5 backbone (PaliGemmaWithExpertModel, PI05Pytorch, helpers)
into pi052/pi05_backbone.py (verbatim copy, no PI05Policy).
- Flatten PI052Policy to subclass PreTrainedPolicy directly (no longer
PI05Policy); inline the needed PI05Policy methods.
- Restore optimizer_foreach/fused + get_optimizer_preset on PI052Config.
- Revert pi0, pi0_fast, pi05 modeling and configuration_pi05 to origin/main
(byte-identical), so the shared policies carry no smolvla modifications.
Behavior verified bit-exact on pepijn223/pi052_robocasa_full: embed_language_
tokens, predict_action_chunk, and the fused flow+text+FAST training loss are
identical before/after (max_abs_diff=0). pi052 tests pass (pre-existing
stale-name collection errors unchanged).
Co-authored-by: Cursor <cursoragent@cursor.com>
* first commit
* feat(policies): add VLA-JEPA
* feat(policies): add VLA-JEPA
* support vla_jepa
* (feat)policies: add VLA-JEPA
* linting
* adding deps to pyproject.toml
* updating uv lock
* adding guards to avoid needing transformers and diffusers for type checking and basic tests
* fixing action and state dim
* fix warnings with qwen processor kwargs
* fixing wm_loss not propagating
* adjusting obs steps, tublets size to match original implementation
* some more fixes to be closer to the original implem
* adding more tests to ensure good coverage
* align VLA-JEPA architecture with original checkpoint
- Remove stale `action_num_heads` / `action_attention_head_dim` config fields;
DiT head dimensions are now always derived from the preset (DiT-B/L/test).
- Add `num_target_vision_tokens` and `action_max_seq_len` config fields required
by the action head's future-token embedding and positional embedding tables.
- Fix default `qwen_model_name` to 2B (matches all released checkpoints).
- Rename `ActionEncoder` attrs w1/w2/w3 → layer1/layer2/layer3 to match
checkpoint key names; replace `nn.Sequential` decoder/state-encoder with
`_MLP2` (layer1/layer2 naming).
- Fix `VLAJEPAActionHead` to size ActionEncoder and StateEncoder at `inner_dim`
(DiT input width) rather than `action_hidden_size` (DiT output width).
- Rename `DiT.blocks` → `transformer_blocks` and `attn` → `attn1` to match
checkpoint; add alternating cross/self attention (even blocks cross-attend to
Qwen context, odd blocks self-attend).
- Add `DiT-test` preset for unit tests.
- Rewrite `ActionConditionedVideoPredictor` with explicit ViT-style blocks
(`_PredictorBlock` with fused qkv) to match checkpoint structure; rename
`encoder`/`norm`/`proj` → `predictor_blocks`/`predictor_norm`/`predictor_proj`.
* propagate action_is_pad masking through VLA-JEPA policy pipeline
Pass the `action_is_pad` tensor from the batch through to the action head
so padded timesteps are excluded from the flow-matching loss.
* update VLA-JEPA tests for arch changes and action_is_pad
- Switch conftest to use `action_model_type="DiT-test"` now that
`action_num_heads` / `action_attention_head_dim` have been removed.
- Add action_head tests covering fully-padded loss (zero) and equivalence
of action_is_pad=None vs all-zeros mask.
- Remove obsolete `test_native_to_lerobot_wm_only` test.
* add VLA-JEPA documentation
Covers architecture overview, pretrained checkpoints, config reference,
training/eval commands for LIBERO-10, and guidance on fine-tuning for
single-camera datasets.
* add one-shot script to convert ginwind/VLA-JEPA checkpoints to safetensors (will remove once migrated)
* make default params more aligned with paper and pretrained models
- adding possibility of freezing qwen backbone and world model
- added tests for weight loading
* trying out to re-init the action head to avoid pretraining dimension mismatch
* allow different state dim and action dim
* removing missleading future_action_window_size to just use chunk_size
* lots of changes to make existing weights work, need to massively refactor the pre and post processing
* refactoring into using pre and post processor
* pre-commit cleanup
* fixing doc defaults args
Signed-off-by: Maxime Ellerbach <maxime@ellerbach.net>
* adressing dtype zeros issue
* adding guard for diffusers
* fixing training and exal examples
* trying to close success rate gap
* fix qwen norm layer output libero eval is now as expected
* adding instructions for different embodiement + fixing some tests
* smol fix to avoid having default CPU device when training
* fixing misconception about multiview / singleview handling
* removing conversion script
* adding licences
* adding .mdx docs and shortening polivy_vla_jepa_README.md
* removing useless pre-processor
* cleanup
* removing swish in favor of silu
* adding configuration gripper index and threshold
* fixing simlink
---------
Signed-off-by: Maxime Ellerbach <maxime@ellerbach.net>
Co-authored-by: ginwind <ginwind@mail.ustc.edu.cn>
Bring the authoritative annotation pipeline from the annotation branch.
The annotation surface is forced to EXACTLY match feat/language-annotation-
pipeline (the annotation branch is the source of truth for annotation
code), which also removes smolvla's stale copies:
- deleted: steerable_pipeline/vocabulary.py, tests/annotations/test_
vocabulary.py, prompts/module_0_vocabulary.txt, module_1_action_record
.txt, module_3_vqa.txt, module_1_plan.txt, and the old module_* prompt
names (now plan_*/interjections_*/vqa.txt).
- synced: all of src/lerobot/annotations/, lerobot_annotate.py,
examples/annotations/, tests/annotations/, datasets/language.py,
tests/datasets/test_language.py, docs/annotation_pipeline.mdx.
Non-annotation conflicts resolved by union (keeping both branches' intent):
- pyproject.toml: keep smolvla's pi extra (+sentencepiece) and add the
molmoact2 extra from main.
- policies/factory.py: keep both dataset_repo_id (pi052 FAST tokenizer)
and dataset_meta (both are referenced); union the policy-type docstring.
- scripts/lerobot_train.py: keep smolvla's pi052 / use_relative_actions
processor-rebuild block.
- uv.lock: regenerated from the merged pyproject.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Drop the optional structured per-subtask action records — not a feature
we want to ship.
* language.py: remove 'action_record' from CORE_STYLES + PERSISTENT_STYLES
(and the matching assertion in tests/datasets/test_language.py).
* config.py: delete ActionRecordsConfig (verb/grasp vocabularies,
frames_per_subtask, emit_record_row) and the PlanConfig.action_records
field.
* plan_subtasks_memory.py: delete _extract_action_record and the
run_episode block that emitted style='action_record' rows; drop the
now-unused json / to_image_blocks imports.
* remove the plan_action_record.txt prompt.
* run_hf_job.py: drop the action_records comment.
Verified: 40 tests pass; pre-commit (ruff, mypy, bandit) clean.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The pyproject annotations-extra comment still described the removed
vllm/transformers in-process backends ('vllm preferred ... transformers
fallback', '_make_vllm_client'); rewrite it for the openai-only reality
and trim it. Also condense the conftest lazy-import NOTE. Comments only.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Bugs
* validator: don't re-raise on unknown style. The second column_for_style
lookup (used to route persistent vs event) now sits in try/except so an
unknown style is recorded by _check_column_routing and skipped instead
of crashing the whole validation pass.
* general_vqa._target_cameras: when restrict_to_default_camera is set but
the configured camera_key isn't one the provider exposes, warn and fall
back to all cameras instead of returning a phantom key that KeyErrors
deep in frame decode.
* interjections: clamp interjection timestamps to frame_timestamps[0]
rather than a hardcoded 0.0 (datasets can start at non-zero t).
Docs / code drift
* annotation_pipeline.mdx: drop the phantom 'vocabulary discovery / phase
0 / --vocabulary.* / canonical_vocabulary.json' section (none of it
exists); describe the real describe->segment + coverage-stitch flow.
Soften the src/lerobot/tools/ + TOOL_REGISTRY reference to 'not part of
this PR' (matches tools.mdx, which already marks the runtime layer as
not-yet-implemented). Fix the --push_to_hub/--new_repo_id wording. Note
the default is now a single h200. Add a 'Contributing new modules'
section inviting module / prompt / quality contributions.
* executor docstring: six phases, no phantom phase 0.
run_hf_job.py
* add the Apache 2.0 license header (was flagged repeatedly).
* default to a single GPU: flavor=h200, parallel_servers=1, num_gpus=1
(scale to h200x4 noted in the docstring).
* pin the install to @main instead of the feature branch (won't break
after merge).
Naming / cleanup
* rename dest_repo_id -> new_repo_id across config / script / example /
test to match the LeRobot dataset edit tools.
* rename prompt templates module_N_*.txt -> descriptive (plan_*,
interjections_*, vqa.txt) and update every load_prompt() call.
* remove dead _messages_to_prompt (used only by the removed in-process
backends).
* declare _warned_decode_fail (frames) and _warned_no_camera (vqa) as
real init=False dataclass fields instead of getattr monkey-patches.
* scope bandit B607 to the two ffmpeg subprocess.run sites via
'# nosec B607' and drop it from the global skip list.
Tests
* fix stale canned-VLM markers ('ONE realistic interruption' ->
'compact interjection', 'Update the memory' -> 'compressed semantic
memory') and drop the dead 'concise hierarchical PLAN' plan responders
(plan generation is deterministic now) in run_e2e_smoke,
test_pipeline_recipe_render, test_modules.
* run_e2e_smoke now asserts interjection + speech rows are produced so a
stale marker can't silently pass again.
* drop remaining 'PR 1' / 'PR 2' references from test comments / names.
Verified: tests/annotations + tests/datasets/test_language +
tests/scripts/test_lerobot_annotate (31 passed); make-style E2E smoke
(interjections=1 speech_atoms=2); pre-commit (ruff, mypy, bandit,
prettier) clean.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The annotation tests had never actually run in CI (collection failed on
the missing 'datasets' extra); now that they do, three stale assertions
surfaced against the evolved pipeline:
* test_module1_plan_memory_subtask_smoke: the memory canned-responder
marker 'Update the memory' no longer appears in module_1_memory.txt
(now 'compressed semantic memory'), so the stub returned no memory
row and the {subtask,plan,memory} subset check failed. Marker
updated to match the current prompt.
* test_module2_mid_episode_emits_paired_interjection_and_speech: the
interjection marker 'Write ONE interjection' is now 'Write ONE
compact interjection' in module_2_interjection.txt, so 0 interjections
were emitted. Marker updated.
* tests/datasets/test_language.py::test_style_registry_routes_columns:
PERSISTENT_STYLES gained 'action_record' in this PR; add it to the
expected set.
These are test/prompt-marker syncs — no production behavior change.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Fast Pytest Tests failed at COLLECTION in the base '--extra test' tier
with 'ModuleNotFoundError: No module named datasets': tests/annotations/
conftest.py imported the fixture dataset builder (-> lerobot.datasets ->
the HF 'datasets' lib + pandas/pyarrow), which only ship under the
'dataset' extra, so the whole annotations package crashed.
Fix uses the repo's proven module-level guard pattern (see
tests/datasets/test_language.py), NOT a conftest-level importorskip —
verified empirically that pytest.importorskip raised during conftest
*import* is treated as a collection ERROR (exit 1), while module-level
importorskip is a clean SKIP.
* conftest.py: import build_annotation_dataset LAZILY inside the
fixtures so the conftest itself imports cleanly in every tier.
* test_modules / test_validator / test_writer / test_pipeline_recipe_
render: add module-level pytest.importorskip('datasets') +
('pandas') before the pyarrow / lerobot.* imports (# noqa: E402 to
match the existing convention). pyarrow-importing modules place the
guard before the pyarrow import.
* tests/scripts/test_lerobot_annotate.py: same guard (its _push_to_hub
path imports lerobot.datasets).
Result:
- base / hardware / viz tiers (no dataset extra): annotation tests
skip cleanly; the rest of the suite runs -> exit 0.
- dataset tier: datasets present -> guards pass through -> annotation
tests run with the stub VLM. The pipeline modules import only
stdlib + relative + lerobot.datasets (no module-level datatrove /
vllm / openai), so they import fine there.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
VideoFrameProvider derived its default camera and camera list from
meta.camera_keys, which mixes image- and video-stored cameras. The
clip/decode paths read videos/<key>/from_timestamp, which only exists
for video keys, so an image-stored camera sorted first (e.g.
observation.images.wrist) crashed the plan phase with a KeyError.
Restrict the list and default to meta.video_keys. Add a regression test
and point the example job at the dataset's actual video camera. Skip
bandit B607 (ffmpeg/git are intentionally resolved via PATH).
Co-authored-by: Cursor <cursoragent@cursor.com>
* feat(rewards): add TOPReward reward model
* refactor(rewards): clean up TOPReward processor/model
* fix(rewards/topreward): add missing input keys mm_token_type_ids
* fix(rewards/topreward): fix pyproject extra typo and simplify processor (#3653)
Add lerobot[topreward] extra to all in
pyproject.toml, drop the redundant labels arg in scoring, and
collapse the dead-branch shape check in the encoder processor.
* optmize topreward input processing (#3660)
---------
Co-authored-by: Cole <91766445+jcoleharrison@users.noreply.github.com>
Co-authored-by: Haoming Song <haomingsong24@gmail.com>
PR #3145 added YAML support for policy.path but left two bugs:
1. extract_path_fields_from_config only deleted config_data[field] when
no sibling overrides existed. With siblings, the dict stayed in place
and draccus crashed decoding it as PreTrainedConfig (no 'type' key).
Sibling overrides go into _config_yaml_overrides and are applied later
by from_pretrained(), so the field can always be removed.
2. wrap() updated config_path_cli to the cleaned temp file path but
never propagated it to the draccus.parse fallback branch. cli_args
still contained --config_path=<original>, so draccus read the
original YAML with path: still present.
Tests passed because they (a) called extract_path_fields_from_config
directly and (b) included type: alongside path: in the YAML, sidestepping
both bugs.
Co-authored-by: Steven Palma <imstevenpmwork@ieee.org>
Replaces the per-layer ``modeling_gemma.eager_attention_forward`` call
with ``torch.nn.functional.scaled_dot_product_attention`` in
``compute_layer_complete`` (pi05) and ``_compute_layer_ki`` (pi052).
PyTorch SDPA picks the memory-efficient kernel for the
block-bidirectional 4D additive mask the dual-expert model uses (FA2 /
FA3 reject it because they only accept causal / sliding-window / varlen
patterns). The shared ``sdpa_attention_forward`` helper mirrors the
eager signature so the call sites are unchanged.
Selective AC: removes the redundant outer ``_apply_checkpoint(forward_func, ...)``
wrap in ``PI05Pytorch.forward``. Per-layer checkpointing inside
``PaliGemmaWithExpertModel.forward`` already handles activation
recompute; the outer wrap was double-recomputing the whole backbone.
+14% steps/sec on its own (job 22161405 vs 22161398, 1xH100).
groot: drop ``@strict`` on ``GR00TN15Config`` — newer ``huggingface_hub``
rejects ``@strict`` on non-dataclass ``PretrainedConfig`` subclasses,
which was blocking imports of any sibling policy through
``lerobot.policies.factory``.
New ``examples/benchmark/bench_pi052_step.py`` (+ slurm sweeps v1..v8)
times PI052Policy.forward+backward (optionally with AdamW) on
synthetic inputs. Headline numbers on 1xH100 with KI=True, GC=True,
L=512, 4.14 B trainable params, AdamW state in bf16:
pre-SDPA eager BS=8 610ms 19.5 GiB -> 13.1 samples/s
sdpa BS=8 + compile=default 413ms 19.5 GiB -> 19.3 samples/s
sdpa BS=16 + compile=default 715ms 37.3 GiB -> 22.4 samples/s
sdpa BS=32 + compile=default 1325ms 44.8 GiB -> 24.2 samples/s
sdpa BS=40 + compile=default 1665ms 48.6 GiB -> 24.0 samples/s
Parity tests in ``tests/policies/pi052/test_pi052_sdpa_attention.py``
cover fp32 / bf16 / GQA / MHA forward + backward — output and grads
match the eager path within bf16 tolerance.
Also ships ``examples/benchmark/fsdp_pi052.yaml`` (FSDP2 accelerate
config wrapping GemmaDecoderLayer + SiglipEncoderLayer) for the
follow-up multi-GPU memory sharding work.
Co-authored-by: Cursor <cursoragent@cursor.com>
Resolves conflicts from 32 commits on main:
* docs/source/_toctree.yml — keep both new toc entries
(annotation_pipeline + video_encoding_parameters).
* docs/source/language_and_recipes.mdx — adopt main's section
ordering (Layer 2 before "Temporal semantics") and float32
timestamp dtype to match the codebase.
* src/lerobot/configs/__init__.py — keep both export sets
(recipe + video encoder).
* src/lerobot/datasets/dataset_metadata.py — drop redundant lazy
imports (top-level imports cover both LANGUAGE_COLUMNS and
DEFAULT_TOOLS); adopt main's @tools.setter for info.json
write-back.
* src/lerobot/datasets/feature_utils.py — call the real
validate_feature_language() instead of returning "".
* src/lerobot/datasets/language.py — float32 timestamps to match
pa.float32() used in video_utils.py and the rest of the codebase.
* src/lerobot/datasets/language_render.py — adopt main's
unwrap_scalar() helper (drops two hand-rolled .item()/list
unwrappers); float32 in docstring.
* src/lerobot/processor/render_messages_processor.py — drop
PR-local _scalar() helper, use shared unwrap_scalar().
* tests/datasets/test_language.py — adopt main's new float32 dtype
+ validate_feature_language warning tests.
* tests/datasets/test_dataset_metadata.py — adopt main's new
tools.setter persist/clear tests.
* uv.lock — regenerated cleanly from main's resolver.
90 of 92 touched tests pass. Two pre-existing test failures
(test_module1_plan_memory_subtask_smoke,
test_module2_mid_episode_emits_paired_interjection_and_speech in
tests/annotations/test_modules.py) are unrelated to this merge —
that test file doesn't exist on main, so the failures originate on
the branch and are addressed by the 8 newer fix(annotate) commits
already on origin that will land in a follow-up.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Parallel variant of build_robocasa_composite_seen.py modeled after the
existing slurm_port_shards.py / slurm_aggregate_shards.py pattern.
Two-phase datatrove pipeline:
* Phase 1 DOWNLOAD: tasks=16 (one per RoboCasa composite_seen task),
each worker downloads its assigned tar via RoboCasa's own
download_datasets helper. Network-bound, idempotent.
* Phase 2 AGGREGATE: tasks=1, single worker calls aggregate_datasets
over the 16 extracted directories. Submitted with depends=phase1 so
SLURM only releases it once all 16 downloads succeed.
Reuses the COMPOSITE_SEEN_TASKS list and per-task download/resolve
helpers from the single-machine script via aliased imports — single
source of truth for 'what does it mean to download a composite_seen
task'.
Local (--slurm 0) mode runs the two phases sequentially in-process for
debugging on a workstation.
Usage on SLURM:
uv run python examples/port_datasets/slurm_build_robocasa_composite_seen.py \
--output-dir=/scratch/${USER}/robocasa_composite_seen \
--hub-repo-id=${HF_USER}/robocasa_composite_seen \
--logs-dir=/scratch/${USER}/logs/robocasa \
--partition=cpu --push-to-hub
Prereq: uv sync --extra annotations (pulls datatrove)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
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>
* chore(gr00t): sync with #3606 for fixing gr00t config crash
* fix(pi0&pi05): fix graph break caused by deepcopy of past_key_values in sample_actions
* fix(pi0&pi05): fix frequent recompile caused by compute_layer_complete
* feat(test): add compile test and benchamrk for pi0 and pi05
* feat(test): add comprehensive testing for pi0 and pi05. Including processor, forward, sample action, etc.
The trained model collapsed to spewing 40+ <loc> tokens for *every*
prompt — subtask, memory, anything — because VQA targets were supervised
to *start* with <loc>. With ~25% of all text samples beginning with a
<loc> token, the LM head learned "Assistant: → <loc>" as a strong
attractor; once one loc is emitted, autoregression chains the rest.
Flip the format so every text target — subtask, memory, speech, AND VQA
— starts with a regular word. The model still learns the <loc>
vocabulary for the spatial portion of the answer, but loc can no
longer be the first generation step out of a clean prompt.
Examples:
point : "green box <loc0162><loc0759>"
bbox : "cube <loc0082>…<loc0409>"
multi : "blue <locs> ; yellow <locs>"
The runtime parser (parse_loc_answer) strips loc tokens and uses the
remainder as label, so it's order-tolerant and works under either
format. Old loc-first checkpoints still parse cleanly at inference;
new training will use label-first.
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>
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>
Pepijn