| --- |
| tags: |
| - pi0 |
| - robotics |
| - openpi |
| - vla |
| datasets: |
| - armnet/busybox_multitask |
| --- |
| |
| # pi0_busybox_multitask |
|
|
| Full-component fine-tune of [π0](https://github.com/Physical-Intelligence/openpi) on SO-101 BusyBox multitask. |
|
|
| | | | |
| |---|---| |
| | **Policy** | π0 (`pi05=false`, full-component) | |
| | **Init** | `pi0_base` | |
| | **Dataset** | [armnet/busybox_multitask](https://huggingface.co/datasets/armnet/busybox_multitask) (132 episodes, 27 tasks) | |
| | **Prompt** | `prompt_from_task` | |
| | **Actions** | 6-D, horizon 30, delta joints | |
| | **Cameras** | `top` → `base_0_rgb`, `wrist` → `left_wrist_0_rgb`, `front` → `base_1_rgb` | |
| | **Training** | 1× H100 NVL, batch 8, 20k steps, cosine 2.5e-5 → 2.5e-6, warmup 1k, EMA 0.999 | |
| | **Result** | loss 0.426 → 0.097 (step 19900). ~2 h 55 m. Code `7679e16` | |
| | **W&B** | [3yuw4r3l](https://wandb.ai/pravsels/busybox_multitask_pi0/runs/3yuw4r3l) | |
|
|
| ## Checkpoints |
|
|
| `checkpoint-19999/` is the final step. `train_state/` is not published. |
|
|
| `checkpoint-5000/` · `checkpoint-10000/` · `checkpoint-15000/` · `checkpoint-19999/` |
|
|
| ```python |
| from huggingface_hub import snapshot_download |
| from openpi.policies import policy_config |
| from openpi.training.config import get_config |
| |
| ckpt = snapshot_download("armnet/pi0_busybox_multitask", allow_patterns="checkpoint-19999/**") |
| policy = policy_config.create_trained_policy( |
| get_config("pi0_busybox_multitask"), |
| f"{ckpt}/checkpoint-19999", |
| ) |
| ``` |
|
|
