Fabyo04/act_so101_cube_v3_050

ACT / SO-101 — cube pick-and-place (v3 setup, 51 demonstrations)

ACT policy trained with LeRobot on Fabyo04/so101_cube_v3 (51 teleoperated demonstrations, 24,404 frames, 30 fps, two RGB views).

Part of a data-efficiency study on imitation learning with a ~$300 SO-101 arm.

Training

Policy	ACT (51.6M params, ResNet18 backbone)
Chunk size / action steps	100 / 100
Steps	015000 (run stopped short of the planned 20,000)
Batch size	32
Epochs	~19.7
Seed	42

The weights at the repository root are checkpoint 015000, the one used for the reported evaluation.

Evaluation

2 / 15 successes on in-distribution cube positions (13%, Wilson 95% CI [3.7%, 37.9%]). Training loss converges, but task success does not follow — the analysis attributes this to a geometric perceptual bottleneck: neither camera view captures gripper and object together at the moment of grasping. A wrist camera is the indicated next step.

Limitations

Single task, single object, fixed lighting. Not evaluated out of distribution. Results come from one seed.