so101_pick_place_v3_structured

Structured MuJoCo pick-and-place demonstrations for the SO-101 arm: 4 cube positions × 10 episodes, generated by a scripted inverse-kinematics controller. This is the training set for smolvla_so101_v5_pretrained_structured.

Episodes	40
Frames	48,000 (1,200 per episode)
Positions	(0.10, -0.10), (0.10, +0.10), (0.20, -0.12), (0.20, +0.12)
Task string	"Pick up the cube and place it in the box."
fps	30
Format	LeRobot v3.0
Cameras	observation.images.front (top-down), observation.images.wrist
State / action	6-D (5 arm joints + gripper), radians

Status

Maintained. The current simulation training set.

How it was collected

Not teleoperation. A damped-least-squares IK solver drives the gripper through Cartesian waypoints with the gripper axis pointing straight down, using a one-jaw grasp strategy (the SO-101's fixed finger rests against the near face of the cube; the moving jaw pins it). Only episodes where the scripted controller succeeded were written.

Limitations

• The four positions are the ones the controller could reach reliably. Two further positions at (0.14, ±0.05) were attempted and yielded 0/10, so the middle of the workspace is unrepresented. This is not a designed spatial sample.
• Scripted trajectories: no human variation across episodes.
• One constant language instruction across all 40 episodes.
• Whether per-episode domain randomisation (cube/box/floor colour, lighting) was active in this generation is not recorded.

Nour El Houda El Iamani — College of Computing, Mohammed VI Polytechnic University (UM6P).