Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
frames
int64
109k
143k
episodes
int64
400
400
first_episode
int64
0
0
last_episode
int64
399
399
state_float32_error
float64
0
0
action_float32_error
float64
0
0
depth_error_m
float64
0
0
numeric_frames_checked
int64
109k
143k
all_numeric_features_exact
bool
1 class
batch_shapes
dict
143,201
400
0
399
0
0
0
143,201
true
{ "observation.images.egocentric": [ 2, 3, 240, 320 ], "observation.depth.egocentric_m": [ 2, 60, 80 ], "observation.depth.valid": [ 2, 60, 80 ], "observation.point_cloud.xyz": [ 2, 2048, 3 ], "observation.point_cloud.rgb": [ 2, 2048, 3 ...
109,200
400
0
399
0
0
0
109,200
true
{ "observation.images.egocentric": [ 2, 3, 240, 320 ], "observation.depth.egocentric_m": [ 2, 60, 80 ], "observation.depth.valid": [ 2, 60, 80 ], "observation.point_cloud.xyz": [ 2, 2048, 3 ], "observation.point_cloud.rgb": [ 2, 2048, 3 ...

Bimanual human demonstrations — new tasks (2026-11)

400 physically successful Superdex human demonstrations per task, simulated in MuJoCo:

  • pick_place_bottle_spam: the left hand grasps a bottle with a natural straight-wrist side grasp (thumb opposite the four fingers) and places it upright in a box; the right hand grasps the YCB Spam can (010_potted_meat_can, thumb on one side face, four fingers on the other) and places it in the box.
  • drawer_spam: the left hand opens the drawer by its handle; the right hand places the Spam can upright in the drawer, then pushes the drawer closed.

Randomization lies between our Round 10 and Round 11 settings: object XY ±2 cm, box/drawer target XY ±1.5 cm, initial wrists ±(1.75, 1.75, 1.25) cm and ±(5, 5, 7.5)°, fingers ±0.04 rad. Every episode is a successful human execution that also passed the robot arm/wrist kinematic reachability check (reachability/, audit evidence only, not robot actions).

Original data only: egocentric RGB (320x240), metric depth with validity, 2048-point camera-visible XYZRGB clouds, calibration, human hand state and actions, timestamps, and the full simulator records needed to replay them. No robot action labels and no inpainted robot videos are included (those are method-specific). corpus.json lists the episodes in order; split.json is the shared train/val split.

Layout and loading

tasks/<task>/human/*.h5 (+ per-episode reports), tasks/<task>/human_lerobot/ (official LeRobot v3; lossless RGB, exact strided depth), tasks/<task>/reachability/. python load_dataset.py --root <dir> --task <task> loads and batch-checks a task. checksums.jsonl/COMPLETE.json list every payload hash.

Actions are 68 human wrist/finger targets at 10 Hz observations (action.substeps keeps the three 30 Hz commands); o_t is measured before a_t. Simulator state in the HDF5 is for replay/validation, not a policy observation.

Rendered Superdex hands and YCB objects retain upstream asset rights; no meshes, credentials or checkpoints are included. No real-robot safety or robot task success is implied.

Downloads last month
-