mobilab-udesc/g1_box_into_box_sim_scripted_v2

G1 box into box (MuJoCo sim, scripted, 15 joints), version 2

200 episodes, 161,176 frames (90 minutes), all successful. Episodes are 23 to 32 s long (mean 26.9 s). Every video frame was decoded once with the decoder LeRobot trains with (torchcodec) before upload.

The episodes are simulated. A MuJoCo Unitree G1 (29 DoF, gloved rubber hands, D405 wrist cameras) runs in a real2sim replica of the lab desk and puts a bubble-wrapped box into a cardboard box. Simulator, scripted policy and recorder: MOBILAB-UDESC/g1-mujoco-real2sim.

Task	put the wrapped box inside the cardboard box
State / action	15 joints: 14 arm joints + waist yaw (Unitree_G1_ArmsWaist_3Cam), action = commanded joint targets
Cameras	cam_head, cam_left_wrist, cam_right_wrist, 640 x 480, 30 fps
Format	LeRobot v3.0 (lerobot 0.4.1)
Per-episode scene	scenes.jsonl: box size, strata, desk / box / light / camera values and the plan's numbers, one line per episode

What changed from version 1 (g1_box_into_box_sim_scripted)

Version 1 was tried on the real robot and failed; version 2 was made from what that test showed.

	Version 1	Version 2
Grasp	hands lowered over the box from above	hands come in beside the box, below its top, then close in sideways
Arms and torso	upper arm up to 1.7 cm inside the torso shell	at least 2.5 cm clear in the plan; measured 2.9 cm or more in every episode
Arms and the cardboard box	not checked	at least 1.2 cm clear in the plan; measured 0.7 cm or more in every episode (never touching)
Joint command speed	up to 2.1 rad/s	at most 0.85 rad/s, command at most 0.24 rad ahead of the joint (the robot-side runners clip at 1.0 rad/s and 0.25 rad)
Box size	22.6 x 22.6 x 26.6 cm	2 to 12 % bigger (mean 7.6 %), a different size every 5 episodes
Box placement	8 x 5.5 cm, yaw +-12 deg, upright	about 11 x 26 cm, one of its four sides toward the robot +-10 deg, upright or upside down
Desk relative to robot	fixed	5 cm nearer to 2 cm farther, 5 cm toward the centre to 2 cm away; the two shifts add up to between -5.5 and +1.5 cm
Light	each light x0.75-1.25	dim / normal / bright levels, per-light variation, warm-cool tint
Camera field of view	fixed	+-4 % per camera

Episodes are kept only when the box ends in the cardboard box and the measured motion kept the limits above. To avoid a dataset of "easy" scenes only, every episode was assigned its scene strata beforehand (orientation, facing side, position cell, desk offset, light), each equally often; a scene that failed was replaced by another one of the same strata.

Stratum	Episodes
Upright / upside down	100 / 100
Side toward the robot (0, 1, 2, 3)	50 / 50 / 50 / 50
Box near / middle / far	67 / 67 / 66
Gap to the cardboard box small / medium / large	67 / 67 / 66
Desk nearer / farther	100 / 100
Desk toward the centre / away from it	100 / 100
Light dim / normal / bright	67 / 67 / 66

It took 615 attempts. 399 never started because the planner found no motion that keeps the clearances, 6 ended with the box outside the cardboard box and 1 broke the lead limit; 9 successful episodes were recorded again with their part (three parts that ran out of attempts, one with a damaged video frame). Success per stratum was between 26 % and 51 %.

Limits

• Simulation only. Object poses come from the simulator; the policy that generated the motions is scripted.
• The robot's arms are short for this task: the box cannot stand farther than 44 cm from the robot, and the desk cannot be farther away than in the ranges above.
• Two corners of the desk range are missing. With the desk both nearer and more toward the centre (shifts adding up to less than about -6 cm) no plan kept the left arm off both the torso and the cardboard box: 0 of 76 attempts. With the desk both farther and more to the left the cardboard box is out of reach. Some of the first 70 episodes were recorded before this limit was set; the lowest is at -5.8 cm.
• Within a stratum the kept scenes still lean toward the easier ones (about one attempt in three is kept).
• About 10 % of the camera frames repeat the previous frame (the simulated camera stream is not locked to the 30 Hz recording).
• Against the cardboard walls the wrapped box counts as 2 cm smaller on every side (bubble wrap squashes); its sides are spring-loaded and give about 8 mm in the grip.