kattri15/lerobot-action-convention-audit

LeRobot Hub action-convention audit

this dataset is a per-dataset audit of what the action tensor actually means in 35 popular lerobot-org datasets on the Hugging Face Hub: delta vs velocity vs absolute control, and gripper polarity (0/1, open/closed, positive/negative all get used, inconsistently). the conventions are recovered from the trajectories themselves with an evidence-first tool, ActionABI, and then compared against each dataset's documented convention (config action.names plus the upstream source spec).

if you train or evaluate a generalist policy across these datasets, or port one of them into your own pipeline, this table tells you where the documented convention is trustworthy, where it is silent, and where it plainly contradicts what the trajectories show.

headline findings

• 35 datasets audited (6 pinned baseline + 29 new), 139 documented field-labels scored, 0 contradictions between ActionABI's recovered evidence and documentation on any field of any dataset.
• "OXE" is not one action convention. Of the 15 Open X-Embodiment ports in this set: 9 use delta end-effector or joint control, 3 use velocity control, 1 is nominally absolute but the trajectory evidence says delta (taco_play, see notes), and 2 are undocumented or mismatched at the tensor level (stanford_kuka_multimodal_dataset's doc describes a 4-dim action but the shipped tensor is 7-dim; ucsd_kitchen_dataset doesn't state delta-vs-absolute at all). All 15 OXE ports lost their semantic RLDS field names in the LeRobot v3.0 conversion (action.names is generic motor_0..N), so the config alone can't tell you the convention for any of them.
• gripper polarity has at least 5 mutually incompatible conventions in this set, including a direct flip: ALOHA's joint-gripper channel is 0 = closed, 1 = open; the OXE gripper_closedness_action convention (ur5, jaco_play, roboturk, nyu_door) is the opposite, 1 = close. Two more OXE variants (berkeley_mvp/berkeley_rpt binary state, toto's open_gripper bool) and taco_play's own scale add a 5th and 6th. ActionABI abstains on gripper for every dataset here (gripper channels are not evidence-identifiable from trajectories alone), so this part of the table is a documentation-side finding, not a recovery.
• unique field certifications: 5 of 35 datasets (pusht, pusht-subtask, and the 3 SO-100/101 sets), all of them documentation-correct, zero false positives.

we are not aware of a prior systematic audit of action conventions across the LeRobot Hub at this scale; if one exists, we'd genuinely like to know about it.

files

• dataset_action_convention_audit.csv — one row per dataset (35 rows), columns below.
• gripper_polarity_summary.csv — the 6-bucket gripper polarity breakdown referenced above.

dataset_action_convention_audit.csv schema

column	meaning
dataset	HF Hub repo id, e.g. lerobot/aloha_sim_insertion_scripted
pinned_revision_short_sha	commit the audit pinned (short SHA)
ecosystem	ALOHA / SO-100/101 / gym-pusht / gym-xarm / LIBERO / MetaWorld / OXE-port / OXE/DROID
action_dim, state_dim	tensor dims at the pinned revision
config_names_status	whether the pinned meta/info.json names the action channels: semantic (named), motor_N (generic, convention erased), opaque (single unnamed field), unknown (not reported for the 6 baseline datasets, which predate this column)
documented_action_convention	absolute / delta / velocity / undocumented, from the pinned config plus the upstream spec
actionabi_recovered_outcome	ActionABI's raw outcome label (e.g. unique_absolute, partial_cartesian, absolute_episode_relative_equivalence, report_without_unique_requirement)
target_field_verdict	ActionABI's own verdict vocabulary for the action-target field: agreement / equivalence_consistent / partial_consistent / abstention_consistent / unlabeled
documented_vs_recovered	a 4-way collapse of the verdict above for quick filtering: agree (agreement or equivalence_consistent), partial_agree (partial_consistent), abstained (documentation exists, ActionABI made no unique claim — this is the honest, dominant outcome, not a gap), undocumented (no usable documented value)
cli_min_loss_target	independent corroboration from ActionABI's compiled C++ scorer on square (action_dim == state_dim) datasets only; not_applicable for non-square dims, unknown where the report doesn't give a value (the 6 baseline datasets predate this cross-check)
gripper_polarity_convention	which of the named polarity conventions (or no_gripper_channel / documented_other_unclassified_polarity / documentation_dimension_mismatch) this dataset's gripper channel follows
gripper_documentation_quote	the verbatim documentation quote the polarity call is based on
gripper_field_verdict	ActionABI's verdict on the gripper field (unlabeled or abstention_consistent for every row — ActionABI never claims gripper)
notes	flagged cases from the source report: the one gate defect that was found and fixed (berkeley_rpt), the taco_play evidence/documentation tension, the 3 frame (not target) partial-discrepancies, and the 2 tensor-level documentation mismatches

row count: 35. every cell is either quoted/derived from the source report and its machine-readable companion files, or explicitly marked unknown/not_applicable where the report doesn't give a value — nothing here is inferred beyond what ActionABI + the report already state.

how this was produced

ActionABI (MIT license, DOI 10.5281/zenodo.21500715) is a C++20 tool for forensic recovery of undocumented robot action contracts. given timestamped trajectories and a finite declared grammar (target: absolute/delta/velocity; frame; permutation; sign; scale; lag; gripper), it reports which fields the passive evidence identifies, which it can only retain as an equivalence set, and which it must abstain on. documentation is used only as a scoring label after the fact, never as an inference input — the passive gate and the C++ scorer never see it.

for this audit, 35 lerobot-org datasets (6 pinned baseline + 29 new) were pulled at a fixed commit each (trajectory parquet only, ~383 MB total across all 35), scored twice (a Python passive gate plus an independent compiled C++ cross-check on square datasets), and compared field-by-field against two documentation layers: the pinned meta/info.json action.names, and the upstream spec (OXE/RLDS TFDS catalog, ACT/ALOHA, LIBERO, MetaWorld, gym-xarm), each with a verbatim quote and source URL. one gate defect was found and fixed during this run (a false absolute/episode-relative equivalence on berkeley_rpt that contradicted its documented delta target); the fix is described in the notes column and the source report.

full method, per-flag case analysis, and reproduction steps: the source report, reports/lerobot_hub_audit.md in the ActionABI repo, and its machine-readable companions results/hub_audit.json / results/doc_agreement.json / results/hub/<dataset>.json.

• ActionABI: https://github.com/Archerkattri/actionabi
• DOI: https://doi.org/10.5281/zenodo.21500715
• sibling project (online action-space adaptation): https://github.com/Archerkattri/actionshift

honest limits

this is a metadata-consistency audit, not independent hardware ground-truth verification. it checks whether ActionABI's trajectory-recovered outcome agrees with each dataset's documented convention, flags the cases where documentation is silent or self-contradictory, and reports where the evidence disagrees with the label. it does not re-run the robots, does not have privileged access to the original data-collection code, and does not certify that the documentation itself is correct against physical ground truth — only that the recovered evidence and the documentation are or are not consistent with each other.

within that scope: passive under-determination is the dominant outcome by design (115 of 139 field labels in the source report are honest abstentions), ActionABI uniquely identified a field on only 5 of 35 datasets, scale and lag are unlabeled everywhere, frame labels are low/medium confidence and the three frame flags in this set (ur5, libero, libero_10) disagree in direction with each other — the report reads this as passive frame identification being genuinely under-determined, not resolved. delta and velocity are indistinguishable from position data alone under constant sample rate (they differ by a scalar dt an affine fit absorbs), so partial_cartesian rows retain both rather than picking one.

citation

if you use this table, please cite ActionABI:

@software{actionabi,
  author = {Attri, Krishi},
  title = {ActionABI: forensic recovery of undocumented robot action contracts},
  url = {https://github.com/Archerkattri/actionabi},
  doi = {10.5281/zenodo.21500715}
}