Dataset Preview
Duplicate
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed
Error code:   DatasetGenerationError
Exception:    CastError
Message:      Couldn't cast
dir: string
contact_weight: double
scored_frames: int64
accepted: int64
contacts: struct<count: int64, dropped_inside_convex_hull: int64, robot_gap_mm: struct<median: double, p90: do (... 734 chars omitted)
  child 0, count: int64
  child 1, dropped_inside_convex_hull: int64
  child 2, robot_gap_mm: struct<median: double, p90: double>
      child 0, median: double
      child 1, p90: double
  child 3, human_gap_mm: struct<median: double>
      child 0, median: double
  child 4, lost_mm: struct<median: double, p90: double>
      child 0, median: double
      child 1, p90: double
  child 5, realized_within_3mm: double
  child 6, material_gap_median_mm: double
  child 7, pad_angle_deg: struct<median: double, within_45: double>
      child 0, median: double
      child 1, within_45: double
  child 8, by_segment: struct<index1: struct<n: int64, robot_gap_median_mm: double, human_gap_median_mm: double>, index2: s (... 411 chars omitted)
      child 0, index1: struct<n: int64, robot_gap_median_mm: double, human_gap_median_mm: double>
          child 0, n: int64
          child 1, robot_gap_median_mm: double
          child 2, human_gap_median_mm: double
      child 1, index2: struct<n: int64, robot_gap_median_mm: double, human_gap_median_mm: double>
          child 0, n: int64
          child 1, robot_gap_median_mm: double
          child 2, human_gap_median_mm: double
      child 2, middle1: struct<n: int64, robot_gap_median_mm: double, human_gap_median_mm: double>
          child 
...
       child 1, robot_gap_median_mm: double
          child 2, human_gap_median_mm: double
intersections: struct<robot_in_material_max_mm: double, robot_in_material_p95_mm: double, material_in_robot_max_mm: (... 74 chars omitted)
  child 0, robot_in_material_max_mm: double
  child 1, robot_in_material_p95_mm: double
  child 2, material_in_robot_max_mm: double
  child 3, frames_over_2mm: int64
  child 4, frames_checked: int64
  child 5, pen_stride: int64
continuity: struct<joint_speed_p95_rad_s: double, joint_accel_p95_rad_s2: double, motion_limit_ratio_max: double (... 33 chars omitted)
  child 0, joint_speed_p95_rad_s: double
  child 1, joint_accel_p95_rad_s2: double
  child 2, motion_limit_ratio_max: double
  child 3, motion_limit_violations: int64
solve_seconds: double
rows: list<item: struct<frame: int64, body: string, segment: string, robot_gap_mm: double, material_gap_mm (... 68 chars omitted)
  child 0, item: struct<frame: int64, body: string, segment: string, robot_gap_mm: double, material_gap_mm: double, a (... 56 chars omitted)
      child 0, frame: int64
      child 1, body: string
      child 2, segment: string
      child 3, robot_gap_mm: double
      child 4, material_gap_mm: double
      child 5, angle_deg: double
      child 6, human_gap_mm: double
      child 7, lost_mm: double
frames: int64
rigid_objects: bool
objects: list<item: string>
  child 0, item: string
finite_geometry: bool
fixed_hand_shape: bool
soft_collision_penalties: bool
unique_track_ids: bool
to
{'frames': Value('int64'), 'objects': List(Value('string')), 'finite_geometry': Value('bool'), 'rigid_objects': Value('bool'), 'fixed_hand_shape': Value('bool'), 'unique_track_ids': Value('bool'), 'soft_collision_penalties': Value('bool')}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1835, in _prepare_split_single
                  for key, table in generator:
                                    ^^^^^^^^^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 621, in wrapped
                  for item in generator(*args, **kwargs):
                              ~~~~~~~~~^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 361, 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 2381, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              dir: string
              contact_weight: double
              scored_frames: int64
              accepted: int64
              contacts: struct<count: int64, dropped_inside_convex_hull: int64, robot_gap_mm: struct<median: double, p90: do (... 734 chars omitted)
                child 0, count: int64
                child 1, dropped_inside_convex_hull: int64
                child 2, robot_gap_mm: struct<median: double, p90: double>
                    child 0, median: double
                    child 1, p90: double
                child 3, human_gap_mm: struct<median: double>
                    child 0, median: double
                child 4, lost_mm: struct<median: double, p90: double>
                    child 0, median: double
                    child 1, p90: double
                child 5, realized_within_3mm: double
                child 6, material_gap_median_mm: double
                child 7, pad_angle_deg: struct<median: double, within_45: double>
                    child 0, median: double
                    child 1, within_45: double
                child 8, by_segment: struct<index1: struct<n: int64, robot_gap_median_mm: double, human_gap_median_mm: double>, index2: s (... 411 chars omitted)
                    child 0, index1: struct<n: int64, robot_gap_median_mm: double, human_gap_median_mm: double>
                        child 0, n: int64
                        child 1, robot_gap_median_mm: double
                        child 2, human_gap_median_mm: double
                    child 1, index2: struct<n: int64, robot_gap_median_mm: double, human_gap_median_mm: double>
                        child 0, n: int64
                        child 1, robot_gap_median_mm: double
                        child 2, human_gap_median_mm: double
                    child 2, middle1: struct<n: int64, robot_gap_median_mm: double, human_gap_median_mm: double>
                        child 
              ...
                     child 1, robot_gap_median_mm: double
                        child 2, human_gap_median_mm: double
              intersections: struct<robot_in_material_max_mm: double, robot_in_material_p95_mm: double, material_in_robot_max_mm: (... 74 chars omitted)
                child 0, robot_in_material_max_mm: double
                child 1, robot_in_material_p95_mm: double
                child 2, material_in_robot_max_mm: double
                child 3, frames_over_2mm: int64
                child 4, frames_checked: int64
                child 5, pen_stride: int64
              continuity: struct<joint_speed_p95_rad_s: double, joint_accel_p95_rad_s2: double, motion_limit_ratio_max: double (... 33 chars omitted)
                child 0, joint_speed_p95_rad_s: double
                child 1, joint_accel_p95_rad_s2: double
                child 2, motion_limit_ratio_max: double
                child 3, motion_limit_violations: int64
              solve_seconds: double
              rows: list<item: struct<frame: int64, body: string, segment: string, robot_gap_mm: double, material_gap_mm (... 68 chars omitted)
                child 0, item: struct<frame: int64, body: string, segment: string, robot_gap_mm: double, material_gap_mm: double, a (... 56 chars omitted)
                    child 0, frame: int64
                    child 1, body: string
                    child 2, segment: string
                    child 3, robot_gap_mm: double
                    child 4, material_gap_mm: double
                    child 5, angle_deg: double
                    child 6, human_gap_mm: double
                    child 7, lost_mm: double
              frames: int64
              rigid_objects: bool
              objects: list<item: string>
                child 0, item: string
              finite_geometry: bool
              fixed_hand_shape: bool
              soft_collision_penalties: bool
              unique_track_ids: bool
              to
              {'frames': Value('int64'), 'objects': List(Value('string')), 'finite_geometry': Value('bool'), 'rigid_objects': Value('bool'), 'fixed_hand_shape': Value('bool'), 'unique_track_ids': Value('bool'), 'soft_collision_penalties': Value('bool')}
              because column names don't match
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1379, in compute_config_parquet_and_info_response
                  parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
                                                                        ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      builder, max_dataset_size_bytes=max_dataset_size_bytes
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 956, in stream_convert_to_parquet
                  builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
                  ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1702, in _prepare_split
                  for job_id, done, content in self._prepare_split_single(
                                               ~~~~~~~~~~~~~~~~~~~~~~~~~~^
                      gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  ):
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1887, in _prepare_split_single
                  raise DatasetGenerationError("An error occurred while generating the dataset") from e
              datasets.exceptions.DatasetGenerationError: An error occurred while generating the dataset

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.

frames
int64
objects
list
finite_geometry
bool
rigid_objects
bool
fixed_hand_shape
bool
unique_track_ids
bool
soft_collision_penalties
bool
258
[ "pipette" ]
true
true
true
true
true

Pipette tasks

Measured RGB-D, reconstructed human hands and objects (joint fit with occlusion and hand self-collision terms, HOPE contact prediction), and kinematic contact-aware Wuji TopoRetarget results.

Six RealSense D435I episodes: three pipetting recordings (pipette_depth_01/02/03), a pipette pickup, and two tube/cap recordings. Each episode has:

  • Measured RGB-D: color frames as JPEG plus rgb.mp4, and color-aligned uint16 depth PNGs.
  • Reconstruction: a joint hand-object fit and HOPE contact predictions.
  • Retargeting: a kinematic contact-aware Wuji hand trajectory for each hand/object pair.

The previous release (contact-free TopoRetarget on the earlier reconstructions) remains available at revision 74c4acb93e9f90315b60f3ceda4a0bcef4a95b4c.

Tube/cap update (2026-11-17): both tube/cap episodes were reconstructed again with corrected tube segmentation. The SAM3 tube prompts are now placed on several frames, and cap pixels are removed from the tube mask. Previously the tube was lost after frame 44 in tube_cap_rs122319, and the clear tube rack was segmented as the tube in tube_cap_rs122335. With the tube tracked, every tube/cap hand is now retargeted to the tube (previously the cap), using the single convex hull of the tube mesh for collision. The earlier tube/cap reconstruction and cap retargeting remain at revision 16cfaf9ebb0a950b88c8e9963c43ea496c70fb3d. The tube/cap results in experiments/hope_reconstruction_v1/ were computed from that earlier reconstruction.

Human reconstruction previews

Original RGB beside the final saved MANO hands and object meshes in one calibrated camera view. These are human reconstructions, not robot retargeting. Invalid hands are omitted; alignment errors in the saved fit are preserved. The tube/cap previews show the 2026-11-17 reconstruction.

pipette_depth_01_tapir

Watch/download human reconstruction · Provenance. 770 frames, 25.67 seconds at 30 FPS. Valid left/right hand frames: 500/417.

pipette_depth_02_tapir

Watch/download human reconstruction · Provenance. 700 frames, 23.33 seconds at 30 FPS. Valid left/right hand frames: 313/400.

pipette_depth_03_tapir

Watch/download human reconstruction · Provenance. 964 frames, 32.13 seconds at 30 FPS. Valid left/right hand frames: 452/631.

pipette_pickup_01_depth

Watch/download human reconstruction · Provenance. 258 frames, 8.60 seconds at 30 FPS. Valid left/right hand frames: 0/175.

tube_cap_rs122319

Watch/download human reconstruction · Provenance. 262 frames, 8.73 seconds at 30 FPS. Valid left/right hand frames: 0/100.

tube_cap_rs122335

Watch/download human reconstruction · Provenance. 283 frames, 9.43 seconds at 30 FPS. Valid left/right hand frames: 110/157.

HOPE contact predictions

Original RGB with the fitted hand vertices projected and coloured by HOPE contact probability (grey 0, yellow 0.5, orange 0.75, red 1). Next to it, the final reconstruction rendered from the recording camera and from the opposite side, which shows the palms. Grey hands have no HOPE prediction for that frame. The per-vertex probabilities are in reconstruction/hope/<hand>_raw.npz (contact_prob); retargeting uses only the defensible subset listed in each contacts.json.

pipette_depth_01_tapir

Watch/download HOPE contacts · Provenance. Poster frame 240 (most contact). Peak vertices with probability ≥ 0.75: left 205, right 232.

pipette_depth_02_tapir

Watch/download HOPE contacts · Provenance. Poster frame 474 (most contact). Peak vertices with probability ≥ 0.75: left 208, right 187.

pipette_depth_03_tapir

Watch/download HOPE contacts · Provenance. Poster frame 580 (most contact). Peak vertices with probability ≥ 0.75: left 285, right 315.

pipette_pickup_01_depth

Watch/download HOPE contacts · Provenance. Poster frame 114 (most contact). Peak vertices with probability ≥ 0.75: left 0, right 332.

tube_cap_rs122319

Watch/download HOPE contacts · Provenance. Poster frame 56 (most contact). Peak vertices with probability ≥ 0.75: left 0, right 237.

tube_cap_rs122335

Watch/download HOPE contacts · Provenance. Poster frame 46 (most contact). Peak vertices with probability ≥ 0.75: left 131, right 162.

Combined robot retargeting previews

Each preview shows original RGB next to all available accepted Wuji hands in one camera-coordinate scene, with the reconstructed objects. Cyan = left hand; coral = right hand. Single-hand recordings show only the available hand. Invalid or rejected frames are labeled and omitted. These independently optimized kinematic hands have not been jointly collision-checked or dynamically validated.

pipette_depth_01_tapir

Watch/download combined retargeting · Provenance. 770 frames at 30 FPS; accepted hand frames: {'left': 491, 'right': 417}; both hands accepted together: 417.

pipette_depth_02_tapir

Watch/download combined retargeting · Provenance. 700 frames at 30 FPS; accepted hand frames: {'left': 310, 'right': 399}; both hands accepted together: 267.

pipette_depth_03_tapir

Watch/download combined retargeting · Provenance. 964 frames at 30 FPS; accepted hand frames: {'left': 449, 'right': 627}; both hands accepted together: 404.

pipette_pickup_01_depth

Watch/download combined retargeting · Provenance. 258 frames at 30 FPS; accepted hand frames: {'right': 175}; both hands accepted together: 0.

tube_cap_rs122319

Watch/download combined retargeting · Provenance. 262 frames at 30 FPS; accepted hand frames: {'right': 98}; both hands accepted together: 0.

tube_cap_rs122335

Watch/download combined retargeting · Provenance. 283 frames at 30 FPS; accepted hand frames: {'left': 109, 'right': 157}; both hands accepted together: 109.

Layout

dataset.json                 per-episode metadata, accepted-frame counts, field descriptions
manifest.json                path, size and sha256 of every file
decompositions/<object>/     CoACD convex parts used for robot-object collision
episodes/<episode>/
  reconstruction/
    calibration.json rgb/ depth/ rgb.mp4 masks/<label>/
    <object>/mesh_local.ply  object mesh (tube/cap also collision.ply: watertight proxy used as the material surface)
    joint_tapir/joint_fit.npz  final hands + objects (camera frame)
    hope/<hand>_raw.npz        HOPE per-vertex contact probabilities (contact_prob) for the fitted hands
    hope/<hand>_recon.npz      MANO parameters HOPE was run on
    hope/contacts.mp4          contact predictions rendered on the reconstruction
    hands_mano.npz             HaWoR hand initialization
    <object>/poses.npz         independent object tracker that initialized the fit (pickup, tube/cap)
  retargeting/<hand>_<object>/
    trajectory.npz scene.xml assets/ retargeted.mp4 summary.json frames.json motion_metrics.json input/
    contacts.json evaluation.json source_frame_ids.npy import_provenance.json

Accepted retargeted frames

Episode Hand / object Source frames Valid hand frames Accepted frames Object collision
pipette_depth_01_tapir left / tube 0–769 500 491 decompositions/tube
pipette_depth_01_tapir right / pipette 0–769 417 417 decompositions/pipette
pipette_depth_02_tapir left / tube 0–699 313 310 decompositions/tube
pipette_depth_02_tapir right / pipette 0–699 400 399 decompositions/pipette
pipette_depth_03_tapir left / tube 0–963 452 449 decompositions/tube
pipette_depth_03_tapir right / pipette 0–963 631 627 decompositions/pipette
pipette_pickup_01_depth right / pipette 0–257 175 175 decompositions/pipette
tube_cap_rs122319 right / tube 0–261 100 86 single convex hull of the tube mesh
tube_cap_rs122335 left / tube 0–282 110 110 single convex hull of the tube mesh
tube_cap_rs122335 right / tube 0–282 157 156 single convex hull of the tube mesh

Conventions

  • Depth: Color-aligned uint16 PNG. depth_metres = pixel * calibration.json depth_scale_m. Zero is invalid. Values are preserved without filtering.
  • Timing: RGB/depth frame filenames share zero-based frame indices. calibration.json includes original color/depth timestamps in milliseconds. Retargeting uses frame_id / 30 seconds; trajectory frame_ids are re-indexed from 0; source_frame_ids.npy next to contacts.json maps them to source frames.
  • Reconstruction: OpenCV camera frame: +X right, +Y down, +Z forward; positions in metres. The final reconstruction for every episode is reconstruction/joint_tapir/joint_fit.npz: left/right MANO joints and vertices, *_valid masks, and _T mapping /mesh_local.ply into the camera frame. The fit uses occlusion and amodal-depth terms (label/free/amodal weights 10/10/4) and a hand self-collision penalty (weight 4, every 4 iterations). hope/_raw.npz holds the HOPE per-vertex contact predictions for the fitted hands (contact_prob, frames x 778 MANO vertices, rows aligned with hope/provenance.json frame_ids; valid marks predicted frames); hope/_recon.npz holds the MANO parameters HOPE was run on. hope/contacts.mp4 shows the predictions on the reconstruction. /poses.npz (where present) is the independent object tracker that initialized the fit, not the final object trajectory.
  • Trajectories: retargeting/_/trajectory.npz: q is N x 20 finger angles in radians ordered by joint_names; wrist_pose and object_pose are N x 4 x 4 in that run world frame; qpos is N x 26 for scene.xml; valid marks solver acceptance; frame_ids index the source frames listed in dataset.json source_frames. Invalid rows must not be treated as targets.
  • Retargeting: Contact-aware TopoRetarget, one run per hand/object: --continuity-experiment limited --seed 0 --max-iter 200 --base-prior-frame source-wrist --contact-weight 10 --contact-fallback --fallback-partial. Defensible HOPE contacts are pulled with weight 10. A frame whose contact solve fails is re-solved without contacts; if that succeeds, the contact solve is retried with a weaker pull, or the pose is moved part-way toward the contact solution, accepting up to 2.5 mm penetration. A frame still rejected starts a new segment, as after an invalid source frame. Collision uses the convex decompositions in decompositions/ unless object_collision says otherwise (the tube/cap episodes use the single convex hull of their tube mesh). contacts.json lists the exported contacts and their filters; evaluation.json has contact realization, penetration and continuity metrics; motion_metrics.json audits velocity/acceleration limits.
  • Continuity: All runs use previous-final-solution initialization, independently scaled wrist/joint continuity and world-frame velocity/acceleration limits within consecutive valid source segments. The earlier contact-free runs remain in dataset revision 74c4acb93e9f90315b60f3ceda4a0bcef4a95b4c; the comparison variants (contact weight 0, contact pull without fallback, fallback without partial progress) in revision a1e1916a0fae7eb5bd8dc2b044b2cd02a883623d.
  • Registration: Each hand was retargeted independently. Each trajectory stores its own T_world_camera; do not assume left/right world frames coincide. These are placement transforms, not measured camera-to-robot calibration or gravity estimates.
  • Physics: Kinematic reference only: no physics optimization, gravity rollout, or PPO. Acceptance is not evidence of a stable physical grasp.
  • Replay: From the g1-wuji-retargeting checkout: ./view_retarget.sh /retargeting/_. Scenes include referenced meshes in assets/ with relative paths.
  • Provenance: Original command.json and source metadata retain historical local paths for provenance; replay scene paths are portable. code_hashes.json records implementation hashes.
  • Code

    License

    Not selected; dataset and bundled third-party asset licensing must be specified before public release.

Downloads last month
286