TaskFocus ablation checkpoints (pi0.5, LIBERO)
pi0.5 (openpi PyTorch) fine-tuned on LIBERO with TaskFocus focus tokens: learned prefix tokens supervised to predict where the gripper, the manipulated object and the target are (now and 10 frames ahead), plus gripper–object relations, contact and task phase. The focus-loss gradient into the backbone is scaled by 0.1. Training: 8 demonstrations per task (all 40 LIBERO tasks), 30k steps, batch 16, cosine LR 5e-5 → 5e-6.
These are ablation checkpoints. Each folder <config>/<experiment>/ holds model.safetensors,
metadata.pt and assets/ (norm stats); configs are pi05_libero_tf_* in the TaskFocus openpi fork.
| Folder (config) | Variant | Seed | LIBERO-10 (50 trials) | LIBERO-Plus layout (150 tasks) | Swapped-instruction success |
|---|---|---|---|---|---|
pi05_libero_tf_focus_obj_low |
where-supervision for object + target only | 43 | 86.4 | 77.3 | 48.0 |
pi05_libero_tf_focus_objgrip_low |
+ gripper | 43 | 85.6 | 75.3 | 38.8 |
pi05_libero_tf_focus_gs01_nowrist_low |
full supervision without wrist-camera object labels | 42 | 83.4 | 73.3 | – |
pi05_libero_tf_focus_gs01_contact_low |
full supervision + contact-region head | 42 | 86.2 | 76.0 | 54.4 |
pi05_libero_tf_focus_gs01_l1w_low |
full supervision on teacher (SAM2 + gripper events) labels incl. wrist masks | 42 | 88.0 | 79.3 | 46.8 |
Reference (not in this repo): plain pi0.5 fine-tune on the same data 82.4 / 72.7 / 31.8; full TaskFocus 87.0 / 79.3 / 46.8 (means over 2 seeds). Research checkpoints, provided as is.