wb_pointed_chair_pull_push_rgb

Whole-body teleoperation data from a Unitree_G1_WholeBody_RGB, published in LeRobot v2.1 format.

Published in the v2.1 layout (one parquet and one video clip per episode) so it loads directly on older lerobot releases. On lerobot v3.0+ run the official upgrade first:

python -m lerobot.datasets.v30.convert_dataset_v21_to_v30 --repo-id=DaoyuanZhu/wb_pointed_chair_pull_push_rgb

Task — pull out the chair indicated by the human gesture, then push it back when the human starts sitting down

At a glance

Episodes	80
Frames	74293 (41 min @ 30 fps)
Episode length	715–1380 frames (median 908)
Cameras	2 × 640×480 H264
State / action	43-D / 43-D joint vectors
Split	train 0:80

⚠️ Two variants — pick the right branch

The dataset ships as two branches that differ only in the dexterous-hand dimensions. Every other column (body joints, timestamps, indices) is byte-identical.

Branch	Hand state (dims 22–28 / 36–42)	Hand action	Use it when
main	all zeros — never recorded	raw Dex3 teleop vector, includes negative values	you only need body/locomotion
real_inspair	Inspire estimate, 0.50–1.00	Inspire native command 0.50–1.00, plus a binary open/close trigger in the last slot	you need anything hand-related

If you train on grasping, use real_inspair. On main the hand state columns are identically zero, so a model will silently learn nothing from them.

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("DaoyuanZhu/wb_pointed_chair_pull_push_rgb", revision="real_inspair")

Schema

observation.state and action are both 43-D, ordered as:

 0–11  legs      left/right  hip_pitch, hip_roll, hip_yaw, knee, ankle_pitch, ankle_roll
12–14  waist     yaw, roll, pitch
15–21  left arm  shoulder_pitch/roll/yaw, elbow, wrist_roll/pitch/yaw
22–28  left hand index_0, index_1, middle_0, middle_1, thumb_0, thumb_1, thumb_2
29–35  right arm shoulder_pitch/roll/yaw, elbow, wrist_roll/pitch/yaw
36–42  right hand index_0, index_1, middle_0, middle_1, thumb_0, thumb_1, thumb_2

In real_inspair, the Inspire hand reports a 6-D estimate: slots 22–27 / 36–41 carry values and slot 28 / 42 stays zero-padded. On the action side the same slots carry the native command and the last slot is a binary trigger (0.0 / 1.0).

The hand state on real_inspair is estimated, transferred from captures that did record Inspire feedback — the legacy rig had no hand sensing. Describe it as estimated/transferred Inspire hand state, not as recorded ground truth.

Teleoperation stream

Beyond the 43-D joint state and action, every frame carries the raw teleoperation and controller signals the capture recorded:

Column	Shape	Content
action.motion_token	(64,)	SONIC motion token
observation.eef_state	(14,)	end-effector: left/right wrist position and quaternion
observation.projected_gravity	(3,)	gravity in the base frame
observation.root_position	(3,)	base position
observation.root_orientation	(4,)	base orientation quaternion
observation.init_base_quat	(4,)	base quaternion at episode start
observation.cpp_rotation_offset	(4,)	controller rotation offset
teleop.left_hand_joints	(7,)	raw left-hand teleop joints (Dex3 order)
teleop.right_hand_joints	(7,)	raw right-hand teleop joints
teleop.left_wrist_joints	(3,)	left wrist roll/pitch/yaw
teleop.right_wrist_joints	(3,)	right wrist roll/pitch/yaw
teleop.smpl_joints	(72,)	SMPL joint positions of the operator
teleop.smpl_pose	(63,)	SMPL pose parameters
teleop.body_quat_w	(4,)	operator body quaternion
teleop.target_body_orientation	(6,)	target body orientation (6-D)
teleop.vr_3pt_position	(9,)	VR three-point tracker positions
teleop.vr_3pt_orientation	(18,)	VR three-point tracker orientations
teleop.planner_mode	(1,)	locomotion mode
teleop.planner_movement	(3,)	commanded movement vector
teleop.planner_facing	(3,)	commanded facing vector
teleop.planner_speed	(1,)	commanded speed
teleop.planner_height	(1,)	commanded base height
teleop.delta_heading	(1,)	heading delta
teleop.stream_mode	(1,)	teleop stream mode
teleop.smpl_frame_index	(1,)	index into the SMPL stream
sync.t_host_ns	(1,)	host timestamp in nanoseconds

Video

Key	Content	Codec
observation.images.cam_chest	cam_chest — raw camera	h264
observation.images.cam_chest_hand_overlay	cam_chest — MediaPipe hand landmarks	h264
observation.images.cam_chest_pose_overlay	cam_chest — RTMO body skeleton	h264
observation.images.cam_head	cam_head — raw camera	h264
observation.images.cam_head_hand_overlay	cam_head — MediaPipe hand landmarks	h264
observation.images.cam_head_pose_overlay	cam_head — RTMO body skeleton	h264

All video is H.264 and plays in any standard player.

Pose annotations

Keypoints ride along as parquet columns, one row per frame — no separate annotation files.

Column	Source	Content
observation.human_hand.<cam>.landmarks_uv	MediaPipe Hands	21 points per hand, pixel u/v, left slot then right
observation.human_hand.<cam>.landmarks_2d	MediaPipe Hands	the same points, normalised x/y/z
observation.human_hand.<cam>.world_landmarks	MediaPipe Hands	metric-like world x/y/z
observation.human_hand.<cam>.valid / .score	—	left/right presence flag and handedness confidence
observation.human_body.<cam>.landmarks_uv	RTMO rtmo-m_16xb16-600e_body7	COCO-17 body, pixel u/v
observation.human_body.<cam>.landmarks_2d	RTMO	the same points, normalised
observation.human_body.<cam>.scores / .valid	—	per-keypoint confidence and a per-frame flag

Zero-filled coordinates with valid = 0 mean nothing was detected; there are no NaNs.

How the annotations were filtered

Body detections are post-processed per episode: score threshold, duplicate suppression by bounding-box containment (plain IoU misses a sprawling false box that swallows the real one), fragment merging, then a single-identity lock built from motion tracklets merged on a saturation-masked torso-hue histogram. Tracklets that overlap in time are never merged, since co-visible detections cannot be the same person.

Hand candidates are additionally anchored to the central person's wrist — a detection further than half a shoulder-width from that wrist is rejected, which is what keeps hand-shaped props out of the annotations.

Posture annotations — when the human sits down

The task hinges on one event: the robot must push the chair back when the human starts sitting down. Every frame is labelled standing / sitting_down / seated, derived from a synchronised markerless 3D mocap of the same sessions rather than hand-annotated.

file	contents
meta/posture/frame_labels.parquet	one row per frame, all 74 293
meta/posture/episode_{i:06d}.json	events and contiguous segments, with provenance
meta/posture_language/episode_{i:06d}.json	the same timeline in natural language, mirroring meta/gesture_language/
meta/human_posture_processing.json	method, thresholds, per-episode summary

Join on episode_index + frame_index — layout-independent, so it works on both this v2.1 layout and a v3.0 conversion. Columns: posture_id (0/1/2), posture, hip_height_norm (the underlying continuous signal, 1.0 standing and ~0.56 seated), language, robot_intent.

class	frames	share
standing	65 033	87.5 %
sitting_down	4 519	6.1 %
seated	4 741	6.4 %

79 of 80 episodes contain exactly one sit event (median 1.80 s). Episode 34 has none — the human genuinely stays standing. standing_up never occurs; every episode ends while seated.

The mocap is already resampled 1:1 onto this timeline, so no interpolation was involved. Normalised hip height is thresholded with hysteresis (>= 0.95 standing, <= 0.68 seated); the onset threshold was tuned so the label lines up with the demonstrator's own reaction, measured from the robot's wrist-motion burst when pushing the chair.

One caveat worth repeating from meta/posture/README.md: the mocap pipeline's "primary" tracklet is the wrong person in 27 of the 80 episodes, locking onto the teleoperator or a bystander. The subject is re-selected per episode as the tallest well-covered tracklet.

A trained recogniser that predicts these labels online from cam_chest alone (224 K parameters, 680 FPS on one CPU thread), plus the scripts that reproduce both the labels and the model, are at DaoyuanZhu/wb_pointed_chair_pull_push_posture.

Layout

data/chunk-000/episode_000000.parquet             one parquet per episode
videos/chunk-000/<video_key>/episode_000000.mp4   one clip per episode
meta/info.json                                    feature schema
meta/episodes.jsonl                               episode index and lengths
meta/tasks.jsonl                                  task strings
meta/stats.json                                   per-feature statistics
meta/gesture_language/episode_000000.json         hand-gesture events as language
meta/posture/frame_labels.parquet                 per-frame posture labels
meta/posture/episode_000000.json                  posture events and segments
meta/posture_language/episode_000000.json         posture timeline as language
meta/human_posture_processing.json                posture labelling provenance
annotations/<cam>/<model>/                        raw per-frame detector output

Loading

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("DaoyuanZhu/wb_pointed_chair_pull_push_rgb", revision="real_inspair")
print(ds.meta.features.keys())
sample = ds[0]

License

Apache-2.0.