1006labs/so101-table-clean-mujoco

SO-101 table clean (MuJoCo)

Simulated demonstrations for the SO-101 arm in MuJoCo, recorded from a scripted expert for finetuning NVIDIA Isaac GR00T N1.7. LeRobot v2.1 format, real-robot LeRobot units (arm joints in degrees, gripper 0-100), 30 fps. This is the training data of the policy 1006labs/so101-table-clean-gr00t at the same revision tag.

Task

A cylinder, a cube and a sphere (all white) sit on a table next to a fixed tray. Every episode is one of three fixed instructions (no paraphrases):

• box: "pick up the cube and put it in the tray"
• ball: "pick up the sphere and put it in the tray"
• cylinder: "pick up the cylinder and put it in the tray"

A task succeeds when the named object lands in the tray; other objects may already be in it. Tasks chain: "pick up the cube" -> success -> "pick up the sphere" from exactly that state.

Every action label comes from a scripted expert that knows the true object poses (the policy only gets the wrist camera and the joints). Two kinds of episodes:

• Scripted: the expert does the whole task from a start condition (clean table, objects already in the tray, a random arm pose, a knocked-over cylinder).
• DAgger: a trained policy runs the task (or three tasks in a row); at chosen frames the sim state is restored and the expert takes over from there, and only the expert's part is recorded (kept only if the expert succeeds). Takeover points: random (any frame), approach (0.5-1.5 s before the policy first moved the target: about to miss or push it), timeout / fell_off (after a failure started), chain (the start of the next task, where the policy left the last one), return (while the policy is coming back home after a success).

Episodes recorded "back home" go on after the drop: the expert lets go, lifts away and returns to the home pose; the others end the moment the object lands in the tray.

Contents

14413 episodes, 2,043,077 frames (1135.0 min), all expert successes.

Source	What	Episodes	Frames	Target cube / sphere / cylinder	Ends
pick-expert-r10	scripted expert, clean table, arm at home	333	56,230	111 / 110 / 112	back home
pick-midstart-r10	scripted expert, 1-2 other objects already in the tray, arm where the last drop ended or near home	333	66,563	111 / 111 / 111	back home
pick-wild-r10	scripted expert, random arm pose, 0-2 objects already in the tray	333	80,536	111 / 111 / 111	back home
pick-fallen-r10	scripted expert, the cylinder lying on its side (as a policy leaves it after knocking it over)	390	93,405	0 / 0 / 390	back home
pick-dagger-r4	expert takeovers from pick-r2/checkpoint-10000 rollouts (mix; policy success 723/800): random 770, approach 73, timeout 48, fell_off 1	892	88,436	315 / 300 / 277	at the drop
pick-dagger-chain-r4	expert takeovers from pick-r2/checkpoint-10000 rollouts (chain; policy success 21/200): random 455, chain 266, timeout 163, approach 31, fell_off 1	916	159,879	333 / 271 / 312	at the drop
pick-dagger-r5	expert takeovers from pick-r4/checkpoint-2000 rollouts (mix; policy success 558/600): random 588, approach 35, timeout 27, fell_off 6	656	68,850	203 / 210 / 243	at the drop
pick-dagger-chain-r5	expert takeovers from pick-r4/checkpoint-2000 rollouts (chain; policy success 207/300): random 770, chain 483, approach 79, timeout 71, fell_off 5	1408	208,361	429 / 456 / 523	at the drop
pick-dagger-r6	expert takeovers from pick-r5/checkpoint-2000 rollouts (mix; policy success 363/400): random 392, approach 35, timeout 26, fell_off 2	455	68,994	140 / 146 / 169	back home
pick-dagger-chain-r6	expert takeovers from pick-r5/checkpoint-2000 rollouts (chain, returning home between tasks; policy success 192/300): random 740, return 657, chain 460, timeout 89, approach 71, fell_off 3	2020	226,587	628 / 627 / 765	back home
pick-dagger-tipped-r8	expert takeovers from pick-r6/checkpoint-2000 rollouts (mix; policy success 1277/1500): tipped 344, approach 211, timeout 201, fell_off 1	757	160,783	0 / 0 / 757	back home
pick-dagger-chain-r10	expert takeovers from pick-r8/checkpoint-2000 rollouts (chain, returning home between tasks; policy success 197/300): random 763, return 688, chain 479, timeout 77, approach 65, tipped 63	2135	253,408	705 / 645 / 785	back home
pick-dagger-tipped-r10	expert takeovers from pick-r8/checkpoint-2000 rollouts (mix; policy success 507/600): tipped 112, approach 90, timeout 81	283	59,533	0 / 0 / 283	back home
pick-dagger-r10	expert takeovers from pick-r8/checkpoint-2000 rollouts (mix; policy success 369/400): random 392, tipped 31, approach 26, timeout 25	474	72,662	151 / 152 / 171	back home
pick-dagger-chain-r11	expert takeovers from pick-r10/checkpoint-2000 rollouts (chain, returning home between tasks; policy success 258/300): random 836, return 816, chain 554, tipped 55, timeout 34, approach 26	2321	260,126	770 / 745 / 806	back home
pick-dagger-tipped-r11	expert takeovers from pick-r10/checkpoint-2000 rollouts (mix; policy success 560/600): tipped 118, approach 39, timeout 33	190	40,307	0 / 0 / 190	back home
pick-dagger-r11	expert takeovers from pick-r10/checkpoint-2000 rollouts (mix; policy success 375/400): random 390, tipped 23, approach 17, timeout 14	444	65,300	151 / 151 / 142	back home
pick-dagger-ball-r14	expert takeovers from pick-r12/checkpoint-6000 rollouts (mix; policy success 955/1000): approach 42, timeout 30, fell_off 1	73	13,117	0 / 73 / 0	back home

• observation.state, action: 6 joints shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper (.pos); the action is the commanded joint target.
• observation.images.wrist: the wrist camera (640x480), the only view of the real robot and the default policy input. observation.images.front: an overhead camera, sim only, for comparison.
• meta/episodes.jsonl adds per episode: seed (object layout, task and start condition), targets, mode (dagger for takeovers), start_placed (objects in the tray before recording), success, source, and for DAgger takeover_reason, takeover_frame, start_mode, stage (task index in a chain).
• meta/sources.json: per source dataset, how it was made (policy checkpoint, seeds, settings).
• meta/modality.json: the GR00T modality layout (the SO100 example's: single_arm 5 + gripper 1).