armanakbari4/fastwam-ur3-drawer-basket

FastWAM UR3 fine-tune (drawer + blue_basket, joint)

Fast-WAM (uncond) fine-tuned on EmbodyX/UR3 real-world tasks drawer_lerobot + blue_basket_lerobot (trained jointly), initialized weights-only from the RoboCOIN-pretrained Fast-WAM. 4000 steps, lr 1e-4 cosine, AdamW(0.9,0.95), bf16, 3 cams (top + L/R wrist, 240x320), 65-frame clips, action_video_freq_ratio 8, action&state dim 14. Tasks:

• drawer: "open the drawer, put the white box inside the drawer then close the drawer"
• blue_basket: "put the medicine then the measuring tape inside the blue basket"

NOTE: loss_action bottomed ~0.037 by step 4000 on only 200 episodes (overfitting) — pick the best checkpoint by REAL-ROBOT success rate, not the last step (2500/3000/3500 often generalize better than 4000). Files: ur3_drawer_basket_step{2500,3000,3500,4000}.pt (weights) + ur3_drawer_basket_dataset_stats.json (norm stats, required for inference).