junn0103/molmoact2-so101-pen-box

MolmoAct2 SO-101 fine-tune (2 tasks, 256 episodes) — step 20,000 total

LoRA-VLM fine-tune of allenai/MolmoAct2 on real SO-101 data (256 episodes / 141,044 frames, single wrist camera, 6-D absolute joint positions). Trained 10k steps from the base checkpoint, then continued 10k more from step 10,000 (~2.3 epochs total at batch 16).

1. "pick the pen and put it in basket" (216 eps)
2. "Place the toy blocks into the cardboard box." (40 eps)

• Recipe: action_mode=both, chunk_size=10, LoRA VLM + fully-trainable action expert, bf16, gradient checkpointing, normalize_gripper=true, quantile normalization, 1x H200
• Root of this repo = step 20,000 (final). Earlier checkpoints under tags step-002500..step-010000; this final also tagged step-020000.
• Curves: first 10k https://wandb.ai/junn0103-seoul-national-university/lerobot/runs/6i93o0sd , continuation https://wandb.ai/junn0103-seoul-national-university/lerobot/runs/jur4pqj5

Rollout: policy.path=this repo, policy.inference_action_mode=continuous, model_dtype=bfloat16.