Zero-WAM ABC bimanual YAM, step 15,000
A research checkpoint fine-tuned from the Zero-WAM pretrained base on a selected, task-balanced subset of LeRobot ABC-130k YAM episodes. This is step 15,000 of the LeRobot training counter: 4 clips per GPU across 8 GPUs and gradient accumulation of 4, for an effective batch of 128 per optimizer update. Video prompt captions were blank. The source dataset is lerobot/abc_130k_v3_train at revision 68651e4929d9fb00f798937b2d62617cab5c771d; the prompt bank contains 12,689 episodes across 201 task labels.
The model is experimental. Physical YAM deployment and task success have not yet been validated. The native model was converted from the distributed training checkpoint; tied weight aliases were checked for equality during conversion. The full distributed optimizer and restart state remain in the original FSx checkpoint and are not included here.
- Downloads last month
- 22