frames int64 54.6k 82.6k | episodes int64 200 200 | first_episode int64 0 0 | last_episode int64 199 199 | state_float32_error float64 0 0 | action_float32_error float64 0 0 | depth_error_m float64 0 0 | numeric_frames_checked int64 54.6k 82.6k | all_numeric_features_exact bool 1
class | batch_shapes dict |
|---|---|---|---|---|---|---|---|---|---|
82,552 | 200 | 0 | 199 | 0 | 0 | 0 | 82,552 | 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
... |
71,724 | 200 | 0 | 199 | 0 | 0 | 0 | 71,724 | 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
... |
63,143 | 200 | 0 | 199 | 0 | 0 | 0 | 63,143 | 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
... |
54,610 | 200 | 0 | 199 | 0 | 0 | 0 | 54,610 | 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
Private research collection: 200 physical Superdex human demonstrations per task:
bottle opening, sequential pick/place, star/triangle dual insertion, and drawer
opening/placement/closing. Read collection.json and COMPLETE.json for current
release status; files may arrive incrementally during upload.
Acceptance means physical HUMAN success plus ROBOT ARM/WRIST reachability.
This release does not establish successful robot contact dynamics. It contains
source human actions, not contact-refined policy labels or inpainted robot videos.
Objects/targets vary within the task configuration; initial hands are identical
within each task. corpus.json fixes accepted IDs and split.json fixes train/val.
Load
Use huggingface_hub.snapshot_download("dexpoint/bimanual-human-demos-800", repo_type="dataset", local_dir="human800") after hf auth login.
Each tasks/<task>/human_lerobot is an independent official LeRobot v3 dataset.
RGB is losslessly embedded in its image feature. Four supplementary inspection
videos are in previews/; they are lossy previews, not the canonical RGB data.
python load_dataset.py --root human800 --task drawer loads synchronized data and
checks a PyTorch batch. Install this project's pinned LeRobot environment (see
requirements-learning-lock.txt) and h5py. No MuJoCo asset checkout is needed to
load observations. The adapter handles relative sidecar paths and float64 clocks.
Channels and timing
RGB320x240, aligned metric optical-axis depth, and2048 visible-surface XYZ/RGB points are recorded at10Hz. Full float32 depth and explicit validity/pixel indices are in the indexed HDF5. LeRobot depth is an exact stride4 subsample, unquantized. The optical camera frame is x right, y down, z forward. Each episode's actual K, world-from-camera transform, clipping and sampling settings are in its source manifest/calibration. Original full depth permits dense backprojection.
o_t is measured before a_t: actions contain68 human wrist/finger targets;
action.substeps preserves all three actual30Hz commands until the next image.
Use simulation.timestamp for the original simulation clock; LeRobot's standard
timestamp is the10Hz dataset clock. Proprioception ordering/units are in
source_manifest.json.schema. Raw numeric channels/integration states remain in
HDF5. Bottle proprioception is reconstructed by MuJoCo forward computation from
recorded integration states, without simulation steps; forces/accelerations in
that adapter are recomputed quantities. Bottle full30Hz integration states,
actions and phase labels are also retained in candidates/*.npz.
Bottle human/*.h5 uses RELATIVE HDF5 virtual datasets referencing
../human_sources/. Download both directories and keep their layout. Other task
HDF5 files are self-contained. Calibration and robot reachability evidence are
privileged metadata, not extra visual policy observations. reachability/ is
kinematic audit evidence and must not be mistaken for physical robot actions.
Verification and rights
Each canonical dataset's validation report checks every numeric frame against
its source, including state/actions, depth, point clouds, timestamps and IDs.
checksums.jsonl records all released payload SHA256 hashes and byte sizes.
Rendered Superdex hands and YCB objects retain upstream asset rights. This private research release grants no additional rights to redistribute underlying assets. No meshes, credentials, training checkpoints or unrelated workspace files are included. Upstream sources: https://github.com/facebookresearch/project_superdex (revision 156fe1394c638a42e6a5f1065a8a70c25f71e292), https://www.ycbbenchmarks.com/, https://github.com/huggingface/lerobot. Consult project asset licenses before redistribution. This collection does not imply real robot safety or physical validation of retargeted robot manipulation.
- Downloads last month
- 12