---
tags:
- pi05
- robotics
- openpi
- vla
datasets:
- villekuosmanen/busybox_multitask
---

# pi05_busybox_multitask

Full-component fine-tune of [π0.5](https://github.com/Physical-Intelligence/openpi) on [`villekuosmanen/busybox_multitask`](https://huggingface.co/datasets/villekuosmanen/busybox_multitask). Relative 6D actions (5 joints delta, gripper absolute) with per-frame task prompts.

This checkpoint **replaces** the 2026-09-01 Hub root. That earlier 30k used shuffled language: OpenPI indexed `meta.tasks` with the inner LeRobot `task_index` instead of the remapped wrapper index. RoboCandyWrapper `main` (`597aa9ad`, remap `de4e4eb`) now maps sample `task_index` onto the sorted `meta.tasks` table, so a `"push the green button"` frame gets that sentence (wrapped index 23), not `"Move the left slider to position 1"`.

Siblings: minmax-norm relative [`pravsels/pi05_busybox_multitask_minmax`](https://huggingface.co/pravsels/pi05_busybox_multitask_minmax); absolute-action [`pravsels/pi05_busybox_multitask_abs`](https://huggingface.co/pravsels/pi05_busybox_multitask_abs) (still the shuffled-language abs run).

| | |
|---|---|
| **Policy** | π0.5 (`pi05=true`, full-component) |
| **Init checkpoint** | `pi05_base` (`gs://openpi-assets/checkpoints/pi05_base`) |
| **Dataset** | [villekuosmanen/busybox_multitask](https://huggingface.co/datasets/villekuosmanen/busybox_multitask) (LeRobot v3, 66 episodes, 12141 frames, 20 fps, 27 tasks) |
| **Prompt** | `prompt_from_task` (RCW remaps `task_index`; 27 instructions; no single default) |
| **Action dim** | 6 (single-arm; 5 joints delta, gripper absolute) |
| **Cameras** | `top` → `base_0_rgb`, `wrist` → `left_wrist_0_rgb`, `front` → `base_1_rgb` |
| **Training** | Vast 4× H100 80GB SXM data-parallel, full-component, 30k steps, global batch 32, action horizon 30 |
| **Norm** | per-timestep 1%/99% quantile |
| **W&B project** | [busybox_multitask_pi05](https://wandb.ai/pravsels/busybox_multitask_pi05) |
| **W&B run** | [4ym0qegc](https://wandb.ai/pravsels/busybox_multitask_pi05/runs/4ym0qegc) |

## Checkpoints

`main` is the latest finalized checkpoint (`params/` + `assets/` at the repo root). The publisher overwrites the root at each 5k save. Training target is step 30,000. `train_state/` is not published.

| Field | Value |
|------|--------|
| Latest published | `main` (`params/`, step 29999) |
| Config | `pi05_busybox_multitask` |
| Code | `84a93ad` on `task/busybox_multitask_promptfix` |
| RCW | git `main` `597aa9ad21176e7f7dcee4aede5dc1ffc07eacee` |
| Runtime | ~4h 30m (2026-09-02 12:40 UTC → 17:10 UTC) |
| Train loss | 0.2669 (step 0) → 0.0030 (step 29900) |

## Usage

```python
from huggingface_hub import snapshot_download
from openpi.policies import policy_config
from openpi.training.config import get_config

ckpt = snapshot_download("pravsels/pi05_busybox_multitask")
policy = policy_config.create_trained_policy(
    get_config("pi05_busybox_multitask"),
    ckpt,
)
```

```bash
uv run scripts/serve_policy.py \
  --policy.config=pi05_busybox_multitask \
  --policy.dir=/path/to/pravsels/pi05_busybox_multitask
```