LeRobot's RoboCasaEnv used a divergent flat state/action layout vs the
robocasa package (robocasa.utils.env_utils.convert_action) and the openpi
robocasa pipeline. This scrambles I/O when using openpi-convention checkpoints
(e.g. the JAX->PyTorch->LeRobot converted pi05 robocasa model: CloseFridge
20% -> 60% once both orders match openpi).
- convert_action: ee_pos(3)+ee_rot(3)+gripper(1)+base_motion(4)+control_mode(1)
- observation.state: ee_pos_rel(3)+ee_rot_rel(4)+base_pos(3)+base_rot(4)+gripper(2)
Matches openpi examples/robocasa/main.py + RobocasaInputs ordering.
Co-authored-by: Cursor <cursoragent@cursor.com>
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>
embed_language_tokens already applies Gemma's sqrt(hidden) normalizer
(GemmaTextScaledWordEmbedding, transformers >=5.4.0). pi052 multiplied FAST
action-token and autoregressive subtask-text embeddings by sqrt(emb_dim) on
top of that, double-scaling them (~2048x). Remove the manual scaling so FAST
and text tokens are single-scaled, consistent with the pi05 fix and OpenPI.
Co-authored-by: Cursor <cursoragent@cursor.com>
transformers >=5.4.0 (PR #44432) makes Gemma's embed_tokens a
GemmaTextScaledWordEmbedding that already multiplies token embeddings by
sqrt(hidden_size). The manual `* sqrt(embed_dim)` applied on top therefore
double-scaled text (~2048x instead of ~45x), breaking VLM alignment for
models trained/run on stock transformers. Remove the manual scaling and rely
on embed_tokens' internal normalizer (matches main #3603). Image features
stay raw (un-normalized), as before.
Co-authored-by: Cursor <cursoragent@cursor.com>
OpenPI (pi0 and pi0-FAST) multiplies language token embeddings by
sqrt(embed_dim) — the Gemma embedder normalizer — before the transformer.
LeRobot pi0/pi0_fast omitted it, leaving text tokens ~45x under-scaled
relative to the residual stream (same class of bug as the pi05 image
scaling). pi0: applied in embed_prefix's lang_embed_func. pi0_fast:
applied inside embed_language_tokens so prompt, FAST action tokens, and
autoregressive next-token embeds are all scaled consistently.
Co-authored-by: Cursor <cursoragent@cursor.com>
lerobot/pi05_base was trained in the OpenPI/big_vision regime where image
(soft) tokens are NOT multiplied by the Gemma embedder normalizer
(sqrt(hidden_size)) — only text tokens are. Scaling image features here
over-scaled them ~45x, breaking the pretrained vision-language alignment
and yielding ~0% closed-loop success on RoboCasa across all pi05 runs.
Co-authored-by: Cursor <cursoragent@cursor.com>
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>
The first-token parity check re-tokenized the decoded (stripped) inference
string, so the leading-space SentencePiece variant always mismatched the
training argmax — a false "DIVERGED" alarm. Remove the autoregressive
inference print and parity comparison (and the now-dead per-sample
select_message generation), keeping only the prompt, ground-truth target,
and teacher-forced argmax accuracy.
Co-authored-by: Cursor <cursoragent@cursor.com>
Collapse the remaining multi-line field comments / docstrings in config.py
to single lines (or two where a knob genuinely needs it), keeping the
essential rationale. Comments only — no field or behavior change.
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>
Tighten the remaining multi-line comment blocks in config.py (derive_task,
frames/window, describe_first, action-record/vqa/vlm fields, video_backend,
repo ids, executor) to 1-3 lines each. Also fix a stale path typo
('examples/annotation' -> the docstring now just says HF Jobs). Comments
only — no field or behavior change.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two behavior-preserving simplifications:
* plan_subtasks_memory.run_episode: the task_aug 'axes' and free-form
branches built identical deduped rows via copy-pasted seen/append
loops. Collapse to one branch that picks the variant source, then a
shared _task_aug_rows() helper does the dedup + row build (-~25 LOC).
* writer: _normalize_persistent_row / _normalize_event_row shared the
same camera-validate + struct construction. Extract _normalize_row(),
keeping the exact key order (the parquet struct schema is inferred
from insertion order, so timestamp must stay between style and camera).
docs: 'Which modules run' is now a table giving each module's on/off flag
(--plan.enabled / --interjections.enabled / --vqa.enabled) and what it
turns off.
Verified: 40 tests pass (incl. test_writer struct round-trip); pre-commit 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>
Dead code (defined but never referenced anywhere in src/tests/examples):
* reader.py: keyframe_indices, episode_frame_timestamps, lookup_data_path,
and the now-orphaned gather_data_paths + episode_offsets_per_path
(lookup_data_path was their only caller).
* staging.py: iter_staged_episodes.
* writer.py: normalize_rows_for_writer.
* config.py VlmConfig: json_mode, batch_size, tensor_parallel_size,
gpu_memory_utilization, trust_remote_code — consumed only by the
in-process vllm/transformers backends that were removed; the openai
auto-serve path carries those vLLM flags via serve_command instead.
Kept max_model_len (still used as the serve-command default).
* config.py TaskAugAxesConfig.total property.
Docs: new 'Key options' section in annotation_pipeline.mdx — grouped
tables (dataset in/out, module toggles, --vlm.*, --plan.*, interjections
+ vqa) describing the flags users actually reach for, with defaults.
config.py: compact the verbose field comments + ActionRecordsConfig /
TaskAugAxesConfig docstrings; fix two stale 'verify' references (the
verify pass was removed — it's describe -> segment now) and the stale
'renders record back to subtask text' note (that path was removed).
vlm_client docstring no longer mentions the removed json_mode field.
Verified: tests/annotations + tests/datasets/test_language +
tests/scripts/test_lerobot_annotate (40 passed); pre-commit clean.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Trim the long inline comment blocks (effective_task / task_aug, action
records, plan-boundary rows, plan-update span closing, windowed +
coverage-stitch sections) and the _generate_plan / run_plan_updates
docstrings to a few lines each. No behavior change.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Same flags and rationale, condensed — each plan-module flag now has a
short one/two-line comment instead of a paragraph.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Top-down flow (read episodes → 3 modules fan out → validator → writer →
parquet) with aligned boxes, instead of the cramped bordered version.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Rewrite annotation_pipeline.mdx in plainer, easier-to-read language
(shorter sentences, active voice, a plain-text intro), add an ASCII
'How it fits together' architecture diagram, and remove the
'Reproducibility via seed and prompt hashes' section. Content/links are
preserved; only wording and structure change.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The example already pins '@main'; update the doc step and the script
docstring from 'the branch under test' to 'lerobot (from main)' now that
the pipeline is merging to main.
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>
action_record is in PERSISTENT_STYLES but was missing from CORE_STYLES,
so STYLE_REGISTRY (= CORE_STYLES | EXTENDED_STYLES) didn't contain it and
the PERSISTENT_STYLES | EVENT_ONLY_STYLES <= STYLE_REGISTRY invariant in
test_style_registry_routes_columns failed. Add it to CORE_STYLES so the
registry, the persistent-set, and column_for_style() stay consistent.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The shipped workflow is Hugging Face Jobs (examples/annotations/run_hf_
job.py): it serves the model with vLLM in the vllm/vllm-openai image and
the pipeline talks to it over the OpenAI-compatible API. The in-process
vllm / transformers local backends added surface (and the vllm
one pinned an old torch) without being part of that path, so they're
removed for now.
* vlm_client.make_vlm_client: keep only backend='openai' (+ 'stub'
rejected with the usual guidance). Requesting 'vllm'/'transformers'
now raises a clear 'not supported for now — use the HF Jobs flow'
error. Removed _make_vllm_client and _make_transformers_client.
* config: backend docstring updated (openai-only); default model_id
bumped to Qwen/Qwen3.6-27B to match run_hf_job.
* docs/annotation_pipeline.mdx: remove the '## Running locally'
section; the launcher description now says one vLLM server per GPU
over the OpenAI API, and the 'One Qwen-VL pass' note drops the
'vLLM/transformers fallback' wording.
Tests are unaffected (they construct StubVlmClient directly; nothing
referenced the removed backends).
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 'dataset' tier failed collecting tests/datasets/test_video_
decoder_cache.py with 'Could not load libtorchcodec ... undefined symbol:
torch_dtype_float4_e2m1fn_x2' — a torch/torchcodec ABI mismatch.
Root cause: the annotations extra's vllm hard-pins an older torch
(via xformers/xgrammar -> torch 2.8). uv resolves a SINGLE unified lock
across all extras, so vllm capped torch to 2.8 for every tier —
including dataset, whose torchcodec 0.11.1 needs torch 2.11. The
result was torch 2.8 + torchcodec 0.11.1 installed together -> ABI break.
(main has no vllm, so it resolves torch 2.11 + torchcodec 0.11.1 cleanly.)
Fix: remove vllm from the annotations extra. It is not needed by
the shipped workflow — examples/annotations/run_hf_job.py gets vllm from
the vllm/vllm-openai image and talks to it over the OpenAI-compatible
API (--vlm.backend=openai), and vlm_client._make_vllm_client imports vllm
lazily. For the in-process --vlm.backend=vllm path, install vllm
separately (the ImportError now says so).
After the fix uv resolves torch 2.11.0 + torchcodec 0.11.1 (matching
main); uv lock --check is clean. The annotations extra still provides
datasets / transformers / openai.
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>
The docs pointed at src/lerobot/datasets/v30/, which does not exist.
Both scripts actually live in src/lerobot/scripts/:
- convert_dataset_v21_to_v30.py
- augment_dataset_quantile_stats.py
Updated the four references (one python -m module path and three
file-path invocations) to the correct location, matching each
script's own usage docstring.
* fix(train): enable relative action overrides for pretrained processors
Keep pretrained processor pipelines when use_relative_actions is enabled and
apply relative/absolute action processor settings through overrides. Rename the
relative action processor registry key to relative_actions_processor.
* fix(config): reject rename_map without pretrained checkpoint
Fail fast when rename_map is set during fresh initialization, since fresh
configs derive feature names from the current dataset and no rename is applied.
---------
Co-authored-by: Pepijn <138571049+pkooij@users.noreply.github.com>
``fit_fast_tokenizer`` previously called ``LeRobotDataset(repo_id,
episodes=[N])`` per sampled episode, which on v3-format datasets
routes through HF datasets' split lookup and raises ``ValueError:
Instruction "train" corresponds to no data!`` on every episode. On
``pepijn223/robocasa_pretrain_human300_v4`` (32 k episodes) this looped
through 13,293 skipped episodes for ~2.5 h before the NCCL watchdog
killed the run via the 2 h ALLREDUCE timeout (job 22182985).
Switch to reading the ``action`` column directly from the dataset's
``data/chunk-*/file-*.parquet`` shards (same pattern as the audit
scripts). Verified end-to-end on the 32 k-episode dataset: 1000 chunks
collected from 1000 episodes in 70.7 s.
Co-authored-by: Cursor <cursoragent@cursor.com>
Quality-gate fix: ruff-format/markdown prettier hook reflow of the
annotation pipeline doc. No content change.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Quality-gate fixes after the main merge:
* UP037: drop redundant quotes from PlanConfig forward-ref annotations
(action_records / task_aug_axes) — safe under 'from __future__ import
annotations'.
* ruff format applied to config.py, executor.py, general_vqa.py,
plan_subtasks_memory.py, validator.py, lerobot_annotate.py.
No behavior change.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Long episodes no longer get sparse subtasks. Previously a long episode
was subsampled to max_video_frames=32 across its whole duration (~1
frame/4s for a 2-min clip). New opt-in windowing keeps a CONSTANT
frames_per_second density by splitting the episode into fixed-length
windows and running the subtask chain per window.
New PlanConfig.subtask_window_seconds (default 0.0 = off). When > 0 and
the episode is longer than one window:
* episode is split into consecutive [w0, w1] windows of this length
* each window's frames are sampled at frames_per_second (so a 32s
window at 1 fps = 32 frames, filling but not exceeding the per-call
context budget)
* the full describe -> segment -> verify chain runs PER window, in
window-relative time [0, L]; spans are offset back to absolute
* all windows' spans are merged, frame-snap-deduped, and stitched into
one contiguous whole-episode cover
Implementation:
* _episode_video_block / _video_message / _describe_episode /
_verify_subtasks gain an optional window=(w0,w1); when set they
embed frames sampled in that absolute range at frames_per_second
(video_url path skipped — it's whole-episode).
* _clean_spans gains bounds= (override clamp range, for window-relative
spans) and dedupe= (skip frame-snap until the merged absolute set).
* new _generate_subtasks_windowed + _subtasks_for_window orchestrate
the loop; _generate_subtasks branches to them when window_s > 0.
run_hf_job.py: --plan.subtask_window_seconds=32 (32s windows at 1 fps).
Cost scales with episode length (chain calls × ceil(duration/window)).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Swap the annotation VLM from Qwen3.6-35B-A3B (sparse MoE, ~3B active)
to Qwen3.6-27B (dense, 27B all-active). Per Scale's dense-captioning
study, model capacity is the #1 lever and the dominant failure is
visual grounding — both helped by ~9x more active params. Qwen3.6-27B
is a vision-language model (vision encoder, image + video), same family
so the chat template / video handling / enable_thinking=false flag are
unchanged, and at 27B dense it still fits one H200 per server, so the
two-parallel-server layout (TP=1, one per GPU) is preserved — no
throughput-layout change, just a much stronger model.
Kept: parallel_servers=2, num_gpus=2, max-model-len 32768 (the 32-frame
embedded budget is ~10k tokens, well under), gpu-mem 0.8.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adopt the one prompt technique Scale's dense-captioning study found
reliably positive: targeted, verb-scoped, visually-grounded
disambiguation rules. Their lesson was that such a rule must fire ONLY
on the spatial situation it names (their narrow 'Stack vs Put' rule
helped; an over-broad directional 'Scoop' rule bled into other verbs
and hurt), so each rule here is phrased visually and scoped to one
confusable pair:
* stack-vs-put (on top of an object vs on a surface)
* insert-vs-put (fitted slot vs surface)
* pick-up/retrieve-vs-put (decide by which way the OBJECT moves:
gripper closes + object moves with hand = pick up; gripper opens +
object stays = put — directly targets Scale's dominant
direction-flip failure)
* pour-vs-put (tilt + flow vs untilted move)
This is the highest-confidence, lowest-risk change from the Scale
findings; our pipeline already aligns with their 'avoid' list (no
temporal tokens, no overlays, no fancy sampling, no sequential context
injection, uniform sampling, describe-don't-predict framing).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Switching the plan module to embedded frames (use_video_url=false)
exposed a context overflow: at frames_per_second=2.0 with the old
max_video_frames=128 default, a 480x640 episode embeds ~128 frames ≈
33-39k vision tokens, over the model's 32768 context — every plan call
died with 'Input length exceeds maximum context length' (HTTP 400),
crashing the whole annotation job.
The video_url path never hit this because the server downsampled; the
embedded path sends every sampled frame, so the frame count is a hard
token budget.
Fix:
* config default max_video_frames 128 -> 32 (~8-10k vision tokens,
comfortable headroom for the prompt + describe/verify passes).
Frames are still sampled UNIFORMLY across the whole episode, so
longer episodes are subsampled, not truncated — full temporal
coverage preserved, just coarser density.
* run_hf_job.py: frames_per_second 2.0 -> 1.0, explicit
--plan.max_video_frames=32, with a comment explaining the token
budget and the 'do not raise toward 128 with embedded frames' rule.
Only the plan module embeds the full episode; VQA (1 frame/tick) and
interjections (4-frame window) were never at risk.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Remove the subtask_full_coverage config flag. Stitching subtask spans
into a contiguous full-episode cover is now always applied in
_generate_subtasks — a sparse / gap-ridden subtask timeline is never
desirable for conditioning, so there's no reason to make it optional.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The verify pass prunes subtasks, which could leave the first subtask
starting after t0 or leave gaps between spans — so the subtask timeline
no longer tiled the episode and frames fell through with no active
subtask label.
New deterministic post-step (no VLM call), default on via
PlanConfig.subtask_full_coverage:
* first subtask start pulled back to the episode's first frame t0
(idle / approach before the first labelled action folds into it)
* each subtask end snapped to the next subtask start (gaps closed)
* last subtask end extended to the last frame t_last
Runs after segment + verify in _generate_subtasks. Starts other than
the first are left as the VLM/verify produced them (already frame-
snapped + distinct), so the cover is contiguous and non-overlapping.
Disable with --plan.subtask_full_coverage=false if a consumer wants
sparse subtasks.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Flip PlanConfig.subtask_describe_first and subtask_verify defaults
False -> True. Every subtask annotation now runs the 3-call grounding
+ pruning chain by default, since the single-call path reliably
hallucinates steps from the task text. Costs 2 extra VLM calls/episode;
disable with --plan.subtask_describe_first=false / --plan.subtask_
verify=false on easy datasets where fewer calls matter more than
label fidelity.
run_hf_job.py: drop the now-redundant explicit flags, leave a note that
the chain is default-on and how to opt out.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The single-call 'watch video -> emit subtask JSON' pattern makes the
VLM commit to structured output before reasoning about what it saw, so
it pattern-matches the task text and hallucinates steps. Split it into
an opt-in multi-call chain that grounds first and prunes last.
New PlanConfig flags (both default False -> single-call unchanged):
* subtask_describe_first: a grounding pass narrates ONLY what is
visible in the video (no subtask JSON yet). That description is
injected into the segmentation prompt via a new {observation_block}
placeholder, so the model segments its own grounded observations
instead of the instruction text. +1 VLM call/episode.
* subtask_verify: after segmentation, an adversarial pass re-watches
the video and drops any candidate subtask it cannot see. Can only
PRUNE (never add/rewrite/move) and fails open (keeps un-verified
spans if the call returns nothing). +1 VLM call/episode.
Implementation:
* _generate_subtasks now orchestrates describe -> segment -> verify.
* Factored span cleaning into _clean_spans (shared by segment + verify
outputs); added _describe_episode and _verify_subtasks helpers.
* New prompts module_1_subtask_describe.txt (returns {description})
and module_1_subtask_verify.txt (returns pruned {subtasks}).
* module_1_subtasks.txt gains a {observation_block} slot at the top.
run_hf_job.py enables both for the RoboCasa run (3 VLM calls/episode
for subtasks). Combined with single-camera grounding + the embedded-
frame path, this is the high-quality configuration.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two fixes for 'subtasks describe actions not in the video' plus a way
to focus the whole pipeline on one camera.
ANTI-HALLUCINATION
1. _episode_video_block: when use_video_url is set but clip extraction
fails, FALL BACK to embedded frames instead of returning an empty
block. An empty block left the VLM with zero visual grounding, so
it invented subtasks from the task text alone — the likely root
cause of hallucinated steps. Now logs a warning and embeds frames.
2. module_1_subtasks.txt gains a GROUNDING preamble (overrides all
other rules): label only motion visible in specific frames; never
invent/anticipate/pad; max_steps is a CEILING not a target; atomic
demos may be exactly ONE subtask; the VIDEO is ground truth, not
the instruction text.
SINGLE-CAMERA GROUNDING
* New VqaConfig.restrict_to_default_camera (default False). When True,
the VQA module grounds on only the --vlm.camera_key stream instead
of iterating every camera — matching the plan / interjection
modules, which already use that single camera. Now the whole
pipeline can focus on one view (e.g. observation.images.base).
run_hf_job.py updated:
* use_video_url=false + frames_per_second=2.0 — embed frames directly
(most reliable; no silent text-only failure mode) with dense
grounding.
* vqa.restrict_to_default_camera=true — VQA on the single camera too.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Drops the replace_subtask_text option and the
_render_action_record_to_subtask_text renderer. Action records are now
strictly additive: when action_records.enabled=True the module emits
style='action_record' rows (the typed {verb,object,arm,grasp,dest,
mistake} schema) and NEVER rewrites the subtask text the policy
conditions on.
The render-back-to-text path was the source of corrupted subtasks
(navigation tasks produced 'move stove to stove', manipulation tasks
got spurious 'with left arm using pinch grip' suffixes). Reconstructing
natural-language subtasks from hallucinated structured fields is
inherently fragile, so the capability is removed rather than guarded.
Removed:
* ActionRecordsConfig.replace_subtask_text field
* PlanSubtasksMemoryModule._render_action_record_to_subtask_text
* the span['text'] = canonical_text overwrite in run_episode
Updated docstrings + run_hf_job.py comment accordingly. emit_record_row
(default True) is now the feature's only output.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
pepijn223