README.md

---
title: Try Pi05
emoji: 🖼️
colorFrom: yellow
colorTo: red
sdk: static
pinned: false
license: wtfpl
short_description: 试玩pi0.5里面的4中LEBERO任务
---

# LIBERO single-task demo

Date: 2026-08-25

## Result

- Checkpoint: `pi05_libero`
- Task suite: `libero_spatial`
- Task id: `9`
- BDDL: `pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate.bddl`
- Policy server: existing OpenPI server on `127.0.0.1:8000`
- Result: `done=True` / success
- Rollout: 174 frames, 17.4 seconds at 10 FPS

## Artifacts

- `videos/single_success.mp4`
- `videos/mid_frame.jpg`
- `logs/single_task.log`
- `libero_config/config.yaml`
- `libero_single_task.py`
- `openpi_libero_helpers.py`

## Additional suites

- `libero_object`, task 0 (`pick_up_the_alphabet_soup_and_place_it_in_the_basket.bddl`): `done=True`, 155 frames, `videos/libero_object_success.mp4`.
- `libero_goal`, task 0 (`open_the_middle_drawer_of_the_cabinet.bddl`): `done=True`, 114 frames, `videos/libero_goal_success.mp4`.
- `libero_10`, task 0 and a retry with task 1: both completed the rollout horizon with `done=False`; the latest result is `videos/libero_10_failure.mp4`. Logs are `logs/libero_10.log` and `logs/libero_10_retry.log`.

## Environment notes

The official LIBERO source was downloaded from:
`https://github.com/Lifelong-Robot-Learning/LIBERO`

The existing OpenPI source and checkpoint were reused. No OpenPI source files were edited.
The legacy LIBERO init-state loader calls `torch.load` without `weights_only=False`, which fails with the installed PyTorch 2.6 behavior. The runner handles this in its own code with `weights_only=False` for the trusted official init-state file.

The `datasets` directory is not needed for this inference rollout; BDDL files, init states, and simulation assets are present.