so101_pick_place_batch2

MuJoCo pick-and-place demonstrations for the SO-101 arm, generated by a scripted inverse-kinematics controller. Second batch of the early simulation generation.

792 episodes recorded with per-episode environment randomisation — cube colour, arm colour, floor appearance and scene lighting all vary across episodes, so the set covers a wide range of visual conditions for the same task.

Episodes	792
Frames	110,880 (140 per episode)
fps	30
Format	LeRobot v3.0
Cameras	observation.images.front, observation.images.wrist
State / action	6-D (5 arm joints + gripper)
Randomised	cube colour, arm colour, floor, lighting

Status

Superseded. Kept public because other work may depend on it. Newer fine-tuning runs use so101_pick_place_v3_structured.

Limitations

• Camera viewpoint. The front stream is a side view of the workspace, not the overhead/front-facing framing SmolVLA was pretrained on and that its fine-tuning recipes assume. This is the main reason the set was superseded: later recordings use a top-down camera at pos="0.20 0.0 0.6". Expect a viewpoint mismatch if you fine-tune a VLA on this data directly.

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