USC-PSI-Lab/SIMPLE-decoupled-wbc-MP

SIMPLE decoupled-WBC, MP tasks (re-rendered)

The 5 motion-planning (MP) tasks of USC-PSI-Lab/SIMPLE-decoupled-wbc as one LeRobot v2.1 dataset, with the observations (states, per-joint columns, video) re-recorded by replaying every demo's actions in the current SIMPLE simulator (see Re-rendering).

• Robot: Unitree G1, 50 fps, egocentric 360x640 H.264 video (observation.images.egocentric; yuv420p, keyframe every 10 frames, no B-frames)
• Episodes: 389, frames: 141101, tasks: 5
• states (32-d) / action (36-d): layout in meta/modality.json

Re-rendering

The hand states of the original MP recordings were captured with the old Dex3 hand gains (kp 80/40, torque limits 5/3 N·m), but their actions were later re-targeted for the soft hand gains SIMPLE now uses (kp 5/2.5, kd 1, torque limits 2.45/1.4/0.7 N·m). Replaying the actions therefore gives different hand states from the stored ones. This dataset fixes that mismatch:

• Unchanged from the MP tasks of SIMPLE-decoupled-wbc: actions (bit-identical), episodes, episode order, episode lengths (including the trailing hold rows), tasks and instructions.
• Re-recorded under the current gains: states, observation.* and the videos. Right/left hand states differ by up to ~1.25 rad (99th percentile ~0.5 rad per episode); arm states by < 0.07 rad. Table textures were resampled in some episodes.
• meta/stats.json (and the identical meta/stats_psi0.json) is recomputed over this data; action statistics are unchanged.
• All 389 episodes are kept. In the replay, 373 succeed at some step and 328 are still successful at their last step (MP success is re-evaluated every step). The 16 that never succeed are kept as recorded.
• The replay used SIMPLE's g1_29dof_wholebody_dex3.xml as of 2026-11-06, before the wrist-joint armature change of 2026-11-09.

State / action layout

All joint values are in radians. For i < 31, states[i] and action[i] are the same joint. The names are also in meta/info.json under features.states.names and features.action.names.

dims	action (36-d)	states (32-d)
0-6	left hand joint targets: thumb_0, thumb_1, thumb_2, middle_0, middle_1, index_0, index_1	measured left hand joints, same order
7-13	right hand joint targets: thumb_0, thumb_1, thumb_2, index_0, index_1, middle_0, middle_1	measured right hand joints, same order
14-20	left arm joint targets: shoulder pitch/roll/yaw, elbow, wrist roll/pitch/yaw	measured left arm joints
21-27	right arm joint targets, same order	measured right arm joints
28-30	waist (torso) roll, pitch, yaw targets	last commanded waist roll, pitch, yaw (see below)
31	base height command (m, 0.41-0.75)	previous step's base height command
32	base forward velocity command vx (m/s, ±0.5)	–
33	base lateral velocity command vy (m/s, ±0.5)	–
34	AMO turning flag (0/1)	–
35	target heading yaw (rad, ±π)	–

Dims 0-31 are absolute targets. Dims 32-34 are velocity commands, marked absolute: false in meta/modality.json.

The two hands use different finger orders: left is thumb, middle, index and right is thumb, index, middle. Both the training data and the SIMPLE eval agent use this order. The per-joint column observation.hand_joints uses thumb, index, middle for both hands.

MP (AMO) conventions

All tasks here use the AMO lower-body controller; the eval agent is simple/baselines/psi0.py (Psi0Agent), which builds the model input the same way.

• states[28:31] = the last commanded waist roll, pitch, yaw: action[t-1, 28:31], and 0 at frame 0. The AMO controller does not use measured waist feedback; its policy input is the previous command, so the state carries that value. This holds exactly at every frame of this dataset.
• states[31] = the last base height command, 0.75 at frame 0.
• action[34] = the AMO turning flag, 0 or 1. It plays the role of the yaw-rate command vyaw in the teleop tasks of SIMPLE-decoupled-wbc. Only the tasks that turn use it (BendPickAndPlaceOnSofaMP, PickAndBendPlaceMP).

Data quality

• Joint order is consistent across all tasks (including the left hand) and matches the eval agent.
• Time axis: states[t] is the observation at frame t and action[t] is the command sent at frame t.
• Other per-joint columns: observation.arm_joints (14), observation.leg_joints (12 leg joints, then waist yaw, roll, pitch), observation.prev_torso_rpy (equals states[28:31]) and observation.prev_height.

Trailing padding

Episode lengths are those of SIMPLE-decoupled-wbc, so its trailing padding is included: episodes whose plain replay (every recorded frame, --no-stop-on-success, success judged at the last step) failed there got their last frame held for up to 50 extra rows (base velocity 0). A held run at the end of an episode can be longer than 50 rows when the source recording already ends with a hold (up to 34 rows in PickAndBendPlaceMP, 21 in BendPickAndPlaceOnSofaMP).

Tasks

task_index	source	instruction	episodes	frames
0	G1WholebodyBendPickAndPlaceOnSofaMP-v0	g1 robot must bend and pick up an object using right hand and place it on the sofa.	75	41933
1	G1WholebodyBendPickMP-v0	bend to pick up the cracker box	91	14112
2	G1WholebodyPickAndBendPlaceMP-v0	pick up the cracker box, turn 180, and drop it in the cushion.	49	53458
3	G1WholebodyTabletopGraspMP-v0	grasp the object on the tabletop.	98	11990
4	G1WholebodyTabletopHandoverMP-v0	pick up the object, then hand it to the left hand.	76	19608

task_index is renumbered 0-4 here (1, 3, 11, 14, 15 in SIMPLE-decoupled-wbc). Episodes are numbered contiguously by task in this order; index is a global 0..N-1.