---
license: apache-2.0
tags: [robotics, vision-language-action, pi0.5, libero]
---
# TaskFocus ablation checkpoints, set 1 (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. Training: 8 demonstrations per task (all 40 LIBERO tasks),
30k steps, batch 16, cosine LR 5e-5 → 5e-6, seed 42. Each folder `<config>/<experiment>/` holds
`model.safetensors`, `metadata.pt` and `assets/` (norm stats). See also `giakhuyendihoc/task-focus-ablations`.

| Folder (config) | Variant | LIBERO-10 (50 trials) | LIBERO-Plus layout (150 tasks) |
|---|---|---|---|
| `pi05_libero_tf_focus_obj_low` | where-supervision for object + target only, grad scale 0.1 | 87.0 | 75.3 |
| `pi05_libero_tf_focus_objgrip_low` | + gripper | 87.2 | 80.0 |
| `pi05_libero_tf_tokens_low` | 4 focus tokens with no focus losses (control) | 84.8 | 72.0 |
| `pi05_libero_tf_focus_gs01_l1_low` | full supervision from an event-based teacher (SAM2 + gripper events) | 87.2 | 72.0 |
| `pi05_libero_tf_focus_gs003_low` | full supervision, backbone grad scale 0.03 | 90.4 | 79.3 |
| `pi05_libero_tf_focus_gs03_low` | full supervision, backbone grad scale 0.3 | 87.4 | 74.0 |

Reference (not in this repo): plain pi0.5 fine-tune on the same data 82.4 / 72.7; full TaskFocus (grad scale 0.1)
87.0 / 79.3 (means over 2 seeds). Research checkpoints, provided as is.