FastWAM UR3 fine-tune (drawer + blue_basket + stacking_cubes, 3-task joint)
Fast-WAM (uncond) fine-tuned on EmbodyX/UR3 real-world tasks drawer + blue_basket + stacking_cubes
(trained jointly on all three), 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"
- stacking_cubes: "put the green cube on top of the black cube and put the red cube on top of the green cube"
NOTE: only ~300 real episodes (100/task) so the model overfits fast (loss_action ~0.066 by step 4000)
-> pick the best checkpoint by REAL-ROBOT success rate, not the last step (2500/3000/3500 often
generalize better than 4000).
Files: ur3_3task_step{2500,3000,3500,4000}.pt (weights) + ur3_3task_dataset_stats.json (norm stats, required for inference).
