rooty2020/RB-Y1_WujiHand2_teleop_ego_exo_100_pointtracks

RB-Y1 + WujiHand2 — ego+exo 3D robot & hand point tracks

3D point tracks for 100 teleop episodes of RB-Y1 with WujiHand2 dexterous hands, ego (ZED-M) + exo (ZED 2i) views. Resampled to 15 Hz (22758 frames).

Per episode (episode_XXX/)

file	contents
robot.npz	arm points (T,512,3), robot base frame
hand.npz	WujiHand2 hand points (T,256,3), robot base frame
ego.mp4,exo.mp4	the two views, 15 fps, intra-only (-g 1)
action.parquet	actions @15 Hz
meta.json	ego+exo intrinsics + T_ego_base/T_exo_base, per-frame ego pose

Projection

Points are in the robot base frame. To a camera: p_cam = R @ p_base + t with meta.cameras.<cam>.T_<cam>_base, then u=fx*x/z+cx, v=fy*y/z+cy.

Differences from the usual bw_jh_dataset format (DROID / RoboCasa / trellis)

This dataset intentionally deviates from our standard convention — read before plugging into the usual loader:

aspect	usual bw_jh format	this dataset
gripper/hand file	gripper.npz (2F gripper)	hand.npz (WujiHand2 dexterous hand, 256)
video names	view1.mp4, view2.mp4 (both exterior)	ego.mp4 (head-mounted, moving) + exo.mp4 (external, static)
point frame	per-frame view1 camera frame (cam_from_world, points already in-camera)	robot base frame (project via T_ego_base/T_exo_base in meta)
tar layout	shard tars of <id>/	one tar of episode_XXX/

So it is NOT drop-in for a loader that assumes points are already in the view1 camera frame — apply meta.cameras.<cam>.T_<cam>_base first. Ask if you want a re-export in the exact usual format (view1/view2, gripper.npz, per-frame view1-cam-frame cam_from_world).

Notes

• FK from observation.joint_position through the source meta/robot_model MJCF; cameras from camera_calibration.json.
• EXO projection is exact; EGO has a small far-left residual = the dataset's own manual ego extrinsic (IoU~0.8), not a pipeline issue.
• Original data 30 Hz → shipped 15 Hz (every 2nd frame, consistent across points/videos/actions).
• Source: kdh8156/RB-Y1_WujiHand2_teleop_ego_exo_100.