so101_pick_place_test0

MuJoCo pick-and-place demonstrations for the SO-101 arm: 88 randomly sampled cube positions, one episode each, generated by a scripted inverse-kinematics controller.

Episodes	88
Frames	105,600 (1,200 per episode)
fps	30
Format	LeRobot v3.0
Cameras	observation.images.front, observation.images.wrist

Status

Superseded, and informative about why. One episode per position turned out to be the wrong structure for small-data fine-tuning: restructuring to 4 positions × 10 episodes (see so101_pick_place_v3_structured) was one of two changes that made fine-tuning work.

Limitations

• One episode per position gives the policy no repetition to learn from at any single configuration.

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