USC-PSI-Lab/SIMPLE-decoupled-wbc

SIMPLE decoupled-WBC (merged)

All 17 tasks from SIMPLE-org/data/data-fixed, merged into a single LeRobot v2.1 dataset, with trailing hold frames appended to the episodes whose plain replay failed (see Trailing padding).

• Robot: Unitree G1, 50 fps, egocentric 360x640 H.264 video (libx264 crf 23, yuv420p, keyframe every 10 frames, no B-frames)
• Episodes: 1517, frames: 606329, tasks: 17
• states (32-d) / action (36-d): layout in meta/modality.json

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	waist roll, pitch, yaw: controller-dependent, 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	teleop: base yaw-rate command vyaw (rad/s, ±1); MP: 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 agents use this order. The per-joint column observation.hand_joints uses thumb, index, middle for both hands.

Teleop vs MP tasks

The data comes from two lower-body controllers. Each task keeps its controller's own convention, and the eval agents in SIMPLE build the model input the same way.

	teleop tasks (*Teleop, decoupled WBC)	MP tasks (*MP, AMO)
states[28:31] waist roll, pitch, yaw	measured waist joint positions at the current frame	last commanded waist roll, pitch, yaw: action[t-1, 28:31], and 0 at frame 0
states[31] height	last base height command, 0.74 at frame 0	last base height command, 0.75 at frame 0
action[34]	yaw-rate command vyaw	AMO turning flag (0/1)
eval agent	simple/baselines/psi0_decoupled_wbc.py	simple/baselines/psi0.py (Psi0Agent)

The AMO controller does not use measured waist feedback. Its policy input is the previous command, so the MP state carries that value.

action[34] has a different name in each controller but the same job: it is the turning command. Treat it as one channel, like vyaw. The teleop data uses the full ±1 range. The MP data uses only 0 and 1, and only in the tasks that turn (BendPickAndPlaceOnSofaMP, PickAndBendPlaceMP).

MP tasks: BendPickAndPlaceOnSofaMP, BendPickMP, PickAndBendPlaceMP, TabletopGraspMP, TabletopHandoverMP. All other tasks are teleop.

Data quality

• Joint order is consistent across all tasks (including the left hand) and matches the eval agents.
• Time axis: states[t] is the observation at frame t and action[t] is the command sent at frame t. In the MP tasks, the waist state was corrected so that it follows the AMO convention above at every recorded frame; on appended hold frames (BendPickAndPlaceOnSofaMP and PickAndBendPlaceMP source tails, and the padded MP episodes) the last state is held, so states[28:31] there can differ from action[t-1, 28:31] by up to ~0.05 rad.
• Other per-joint columns: observation.arm_joints (14), observation.leg_joints (12 leg joints, then waist yaw, roll, pitch), observation.prev_torso_rpy and observation.prev_height. In MP tasks, observation.prev_torso_rpy equals states[28:31].

Trailing padding

Episodes whose plain replay (every recorded frame, --no-stop-on-success, success judged at the last step) failed got the last frame held for 50 extra rows (base velocity 0), with the recorded rows unchanged. 600 episodes are padded; the other 917 are not.

Padding is capped at 50 rows per episode. The padding was first appended in up to 4 rounds of 50 (0-200 rows); 312 episodes had more than 50, and their extra rows (39750 in total) were removed from the end of the episode and its video (scripts/cap_trailing_padding.py in SIMPLE; per-episode record in cap_padding_log.jsonl). 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 and BendPickAndPlaceTeleop).

Merge notes

• Episodes are renumbered contiguously, with sources appended in alphabetical order. index is rewritten to a global 0..N-1.
• meta/stats.json (and the identical meta/stats_psi0.json) is recomputed over all frames of the merged data (mean/std/min/max/q01/q99, float64 accumulation). Every row of meta/episodes_stats.jsonl is computed from its full episode.
• Changes since the previous version (1584 episodes / 600021 frames): MP tasks re-targeted for the soft Dex3 hand gains (hand targets action[0:14] changed) with failed replays removed (BendPickAndPlaceOnSofaMP 99→75, BendPickMP 100→91, PickAndBendPlaceMP 57→49, TabletopGraspMP 100→98, TabletopHandoverMP 100→76); trailing padding as above. Teleop recorded rows are unchanged.
• Padding capped at 50 rows (previously up to 200; 1517 episodes / 646079 frames before the cap): only padded rows were removed, index, next.done, meta/episodes.jsonl, meta/episodes_stats.jsonl and meta/stats.json were rewritten.
• Videos re-encoded with a keyframe every 10 frames and no B-frames (-g 10 -keyint_min 10 -sc_threshold 0 -bf 0, SIMPLE c5fb4bd; previously x264 defaults: keyint 250, 3 B-frames), so a random-frame lookup decodes ~5 frames instead of ~95. Re-encoded from the previous H.264 files (PSNR vs previous ≈ 43 dB); frame counts and all parquet/meta files unchanged.

Tasks

task_index	source	instruction	episodes	frames
0	G1WholebodyBendHandoverTeleop-v0	bend the robot and pick up the cracker box, then place it on the container.	96	44575
1	G1WholebodyBendPickAndPlaceOnSofaMP-v0	g1 robot must bend and pick up an object using right hand and place it on the sofa.	75	41933
2	G1WholebodyBendPickAndPlaceTeleop-v0	bend to grasp the cracker box and drop it in the basket.	48	23958
3	G1WholebodyBendPickMP-v0	bend to pick up the cracker box	91	14112
4	G1WholebodyBendPickTeleop-v0	bend the robot and pick up the cracker box	97	22473
5	G1WholebodyCloseDoorTeleop-v0	move forward to the door and close it	99	32251
6	G1WholebodyHandoverTeleop-v0	Hand over cracker box from right hand to left hand and place it on the container.	100	50530
7	G1WholebodyLocomotionPickBetweenTablesTeleop-v0	pick up the cracker box from table1,locomotion to table2,and place on table2.	99	69464
8	G1WholebodyOpenFaucetTeleop-v0	turn the faucet	99	38458
9	G1WholebodyOpenOvenTeleop-v0	move forward to the oven and open it	99	38313
10	G1WholebodyOpenTrashCanTeleop-v0	move forward to the trash can and open it	99	36985
11	G1WholebodyPickAndBendPlaceMP-v0	pick up the cracker box, turn 180, and drop it in the cushion.	49	53458
12	G1WholebodyPickAndPlaceAndHugContainerTeleop-v0	pick up the apple from table1,hug the container,walk to table2,and place on table2.	94	46588
13	G1WholebodyPushOfficeChairTeleop-v0	move forward to the office chair and push it to the table	99	35979
14	G1WholebodyTabletopGraspMP-v0	grasp the object on the tabletop.	98	11990
15	G1WholebodyTabletopHandoverMP-v0	pick up the object, then hand it to the left hand.	76	19608
16	G1WholebodyXMovePickTeleop-v0	move forward to pick up the cracker box	99	25654