---
license: apache-2.0
tags:
  - openpi
  - pi0.5
  - vla
  - flow-matching
  - robotics
  - arx-x5
  - real-robot
language:
  - en
pipeline_tag: robotics
---

# pi0.5 ARX deltapose checkpoint — `puttube` task (step 29999)

OpenPI π0.5 fine-tuned on ARX X5 single-arm real-robot data for the
"Move the test tube from the transparent rack to the yellow wooden rack" task.

- **Base model**: pi0.5
- **Action representation**: deltapose (6-DoF delta xyz+rpy + binary-snap gripper)
- **Action horizon**: 10
- **Train data**: 415 episodes, 97 483 frames, 30 fps
- **Train steps**: 30 000 (cosine LR, peak 5e-5)
- **Format**: Orbax JAX checkpoint (`params/` + `assets/xzj_data_0511/norm_stats.json`)

> Inference-only — `train_state/` (optimizer + EMA) was stripped to keep the
> upload small (~12 GB vs full 42 GB).

## Download

```bash
huggingface-cli download magic0/forcevla-flexiv-tactar-june26 --local-dir ./checkpoints/29999
```

## Deploy

Server code: https://github.com/SII-ZijunX/openpi-inpaint-vla (branch `inpaint-vla`).

```bash
git clone https://github.com/SII-ZijunX/openpi-inpaint-vla.git
cd openpi-inpaint-vla
uv venv && source .venv/bin/activate
uv pip install -e ".[pytorch]"

# IMPORTANT: --asset_id xzj_data_0511 is required because this checkpoint
# stores norm_stats under assets/xzj_data_0511/, not under the original
# absolute training-data path.
python scripts/inference_server_xzj_arx.py \
    --checkpoint_dir ./checkpoints/29999 \
    --asset_id xzj_data_0511 \
    --zmq_port 6789
```

## Internals

- `params/` — Orbax/OCDBT sharded weights, load via
  `openpi.models.model.restore_params(ckpt/"params", dtype=jnp.bfloat16)`
- `assets/xzj_data_0511/norm_stats.json` — z-score / quantile stats from
  training data, required at inference time