Datasets:

sajio
/

soma-to-multiple-robots

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
completed: list<item: struct<iteration: int64, returncode: int64, report: string>>
  child 0, item: struct<iteration: int64, returncode: int64, report: string>
      child 0, iteration: int64
      child 1, returncode: int64
      child 2, report: string
pending: list<item: int64>
  child 0, item: int64
protocol: string
caveat: string
motions: list<item: struct<clip_id: string, episodes: int64, full_clip_completion_rate: double, mean_episode_ (... 118 chars omitted)
  child 0, item: struct<clip_id: string, episodes: int64, full_clip_completion_rate: double, mean_episode_fraction: d (... 106 chars omitted)
      child 0, clip_id: string
      child 1, episodes: int64
      child 2, full_clip_completion_rate: double
      child 3, mean_episode_fraction: double
      child 4, root_relative_body_error_m: double
      child 5, global_anchor_error_m: double
      child 6, reference_contact_match: double
sampling_observations_recorded: double
clip_macro_completion_rate: double
to
{'protocol': Value('string'), 'caveat': Value('string'), 'clip_macro_completion_rate': Value('float64'), 'sampling_observations_recorded': Value('float64'), 'motions': List({'clip_id': Value('string'), 'episodes': Value('int64'), 'full_clip_completion_rate': Value('float64'), 'mean_episode_fraction': Value('float64'), 'root_relative_body_error_m': Value('float64'), 'global_anchor_error_m': Value('float64'), 'reference_contact_match': Value('float64')})}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2840, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              completed: list<item: struct<iteration: int64, returncode: int64, report: string>>
                child 0, item: struct<iteration: int64, returncode: int64, report: string>
                    child 0, iteration: int64
                    child 1, returncode: int64
                    child 2, report: string
              pending: list<item: int64>
                child 0, item: int64
              protocol: string
              caveat: string
              motions: list<item: struct<clip_id: string, episodes: int64, full_clip_completion_rate: double, mean_episode_ (... 118 chars omitted)
                child 0, item: struct<clip_id: string, episodes: int64, full_clip_completion_rate: double, mean_episode_fraction: d (... 106 chars omitted)
                    child 0, clip_id: string
                    child 1, episodes: int64
                    child 2, full_clip_completion_rate: double
                    child 3, mean_episode_fraction: double
                    child 4, root_relative_body_error_m: double
                    child 5, global_anchor_error_m: double
                    child 6, reference_contact_match: double
              sampling_observations_recorded: double
              clip_macro_completion_rate: double
              to
              {'protocol': Value('string'), 'caveat': Value('string'), 'clip_macro_completion_rate': Value('float64'), 'sampling_observations_recorded': Value('float64'), 'motions': List({'clip_id': Value('string'), 'episodes': Value('int64'), 'full_clip_completion_rate': Value('float64'), 'mean_episode_fraction': Value('float64'), 'root_relative_body_error_m': Value('float64'), 'global_anchor_error_m': Value('float64'), 'reference_contact_match': Value('float64')})}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Thank you, Lambda

We thank Lambda for the compute support behind this multi-robot motion-generation and simulation effort.

Please let us know if you have feedback and suggestion! chil@apocynthion.ai / lloyd@apocynthion.ai

SOMA to Multiple Robots

AlphaMotion is our self-developed cross-embodiment motion model. This repository collects robot-specific motion references generated from SOMA, visual comparisons, and saved downstream simulation evidence. It currently covers Unitree H2, Agibot A3, X2, Tiangong, EngineAI T800 and Fourier GR3. AlphaMotion is the model name; these are target robots, not model variants.

Known hand-mapping defect in the T800, X2 and GR3 releases: The current wrist/palm-frame mapping rotates the hands by approximately 90° from the intended outward-facing orientation. This affects the archived reference trajectories, not just their preview videos. Hand pose, wrist-joint trajectories, and end-effector orientation should not be treated as validated or used for manipulation or hardware control without correction and re-evaluation. We are publishing these versions for transparent research review, not representing them as hand-accurate releases. The earlier H2, A3 and Tiangong datasets are separate releases; this warning does not itself establish their hand-orientation quality.

The source material is Motion Data by Bones Studio (BONES-SEED), subject to its separate source terms. This repository does not contain the original SOMA arrays or AlphaMotion weights. The owner will provide the applicable release terms separately; see LICENSE.md.

Robot datasets

Target	Motion files	Rate	Current evidence	Details
Unitree H2	132,788	50 Hz	Format/integrity audit; selected simulation-policy evaluations; no full-corpus physical-QC pass rate	H2 dataset
Agibot A3	132,788	30 Hz	Independent archive verified; 1,770/2,000 (88.5%) passed the frozen geometric sample screen	A3 dataset
Tiangong	132,788	30 Hz	Independent archive verified; 1,810/2,000 (90.5%) passed the frozen geometric sample screen	Tiangong dataset
T800	132,788	30 Hz	Two robot-only archive parts; known hand-mapping defect; full-corpus physical QC not established	T800 dataset
X2	132,788	30 Hz	Two robot-only archive parts; known hand-mapping defect; full-corpus physical QC not established	X2 dataset
GR3	132,788	30 Hz	Two robot-only archive parts; known hand-mapping defect; full-corpus physical QC not established	GR3 dataset

Visual review

The H2, A3 and Tiangong cards contain animated 3-by-3 SOMA/robot comparisons. The new T800, X2 and GR3 cards each contain an animated 2-by-2 comparison in the same UI-free visual style. Geometric near-floor surfaces are pale green; excessive penetration is red. These are qualitative examples, not random-sample QC estimates or force measurements. Their clip provenance is supplied alongside the GIFs; the new galleries visibly retain the known hand-mapping defect.

What the evaluations establish

Each robot is evaluated on distinct levels: file production, archive integrity, frozen kinematic screening, simulation tracking on selected motions, and (only if available) physical hardware testing. The A3/Tiangong screen used the same 2,000 uniformly sampled source IDs and the frozen engineering thresholds of 15 cm/s for 0.1 s support-phase foot material speed and 5 mm for 0.1 s whole-surface floor penetration. It is a geometric screen, not a dynamics or hardware test. A3 had 203 failures and 27 unknown cases; Tiangong had 167 failures and 23 unknown cases. Unknown cases are not counted as passes. See each robot card and the per-clip records for category breakdown and limitations. We do not transfer an H2 QC number to A3 or Tiangong, or claim the three robot cohorts form a directly pooled benchmark.

Saved policy checkpoints and evaluations provide downstream trainability evidence for selected AlphaMotion-generated references, especially H2. They do not certify every motion, imply a generalist policy has succeeded, or constitute a real-robot safety claim. The cohorts used different protocols; their success counts are not one pooled benchmark.

For T800, X2 and GR3, packaging integrity and four-clip visual previews are documented separately from physical QC and policy success. Because their hand mapping is known to be wrong, even a lower-body policy tracking result would not validate their wrist or end-effector motion. No pass rate for these three full releases is claimed here.

Repository layout

data/shards/, data/index.jsonl and data/MANIFEST.json: existing H2 motion release.
data/a3/: separate A3 NPZ archive, manifest and QC evidence when verified.
data/tiangong/: separate Tiangong NPZ archive, manifest and QC evidence when verified.
data/t800/parts/, data/x2/parts/, data/gr3/parts/: two disjoint, independently extractable robot-only NPZ archive parts per target; associated manifests document counts and hashes.
assets/: UI-free animated contact comparisons and provenance.
policies/: selected checkpoints with saved evaluation records.
H2, A3, Tiangong, T800, X2, GR3: per-robot format, evidence and limitations.

These are kinematic reference motions, not measured motor commands. Before control training or hardware use, apply the matching robot model, joint mapping and environment, and screen each motion. Motion quality and policy-learning success are related but not interchangeable.

Downloads last month: 717

Total file size:

57 GB