FastWAM UR3 fine-tune (drawer + blue_basket + stacking_cubes, 3-task joint) — 10k steps
Fast-WAM (uncond) fine-tuned on EmbodyX/UR3 real-world tasks drawer + blue_basket + stacking_cubes
(trained jointly), initialized weights-only from the RoboCOIN-pretrained Fast-WAM.
10,000 steps on 6 GPUs (effective batch 48) — a longer run than the earlier 4k version, since 3 tasks
need more training. 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: ~300 real episodes (100/task) so the model overfits — pick the best checkpoint by REAL-ROBOT
success rate, not the last step. With 10k steps the mid-to-late checkpoints (6k-9k) are the strongest
candidates. Checkpoints provided: step {4000, 5000, 6000, 7000, 8000, 9000, 10000}.
Files: ur3_3task_10k_step{4000..10000}.pt (weights) + ur3_3task_10k_dataset_stats.json (norm stats, required).
