---
library_name: openpi
tags:
- robotics
- imitation-learning
- diffusion-policy
- pi0.5
- robotwin
- arx-x5
---

# pi0.5 — RoboTwin ICL ARX-X5 v4mix — step 20000

Fine-tune of the Physical Intelligence
[pi0.5](https://github.com/Physical-Intelligence/openpi) base policy on the
RoboTwin ICL benchmark for the ARX-X5 dual-arm robot, with additional
paired-variant demonstrations mixed in (v4mix).

## Contents

| File | Purpose |
|---|---|
| `model.safetensors` | ~7.0 GB — main model weights (bf16) |
| `optimizer.pt` | ~13 GB — AdamW state; needed only to resume training |
| `metadata.pt` | Lightning bookkeeping (step, epoch, RNG state) |
| `assets/robotwin-icl-arx-x5-v4mix/norm_stats.json` | Per-dim action/state norm stats used at inference |

## How to load

```python
# Requires openpi (Physical Intelligence). See
# https://github.com/Physical-Intelligence/openpi
from openpi.policies import policy_config
from openpi.training import config as pi_config

train_cfg = pi_config.get_config("pi05_robotwin_icl_arx_x5_v4mix")
policy = policy_config.create_trained_policy(
    train_cfg,
    checkpoint_dir="path/to/this/checkpoint",
)
actions = policy.infer(obs)["actions"]   # (50, 14) action chunk
```

## Training recipe

- **Base**: pi0.5 (openpi `pi05_base` params)
- **Continued from**: `pi05_robotwin_arx5_icl_20260702/200000` (an earlier
  ICL-only fine-tune)
- **Config**: `pi05_robotwin_icl_arx_x5_v4mix`
- **Model**: `Pi0Config(action_dim=32, action_horizon=50, pi05=True,
  paligemma_variant='gemma_2b_lora', action_expert_variant='gemma_300m',
  dtype=bfloat16)`
- **Data**:
  - **Canonical**: 3000 train + 1000 val episodes from
    `robotwin-arx5-lerobot` (25-task ICL split)
  - **Variants**: 1500 train episodes from `robotwin-arx5-variants-lerobot`
    (paired-v4 waypoint / grasp-point / arm variants)
- **Optimizer**: AdamW(b1=0.9, b2=0.95, eps=1e-8, wd=1e-10, clip_grad_norm=1.0)
- **LR schedule**: CosineDecay(warmup=200, peak_lr=3e-5, decay_steps=50000)
- **Freeze filter**: LLM backbone frozen except `_1` (action expert) and
  LoRA adapters; PaliGemma trained with `gemma_2b_lora`
- **This checkpoint**: step **20000 / 50000**, batch size 16 × 4 GPUs (FSDP),
  bfloat16 precision
- **Trained on**: 2026-07-06

## Notes on step choice

Prior top-K noise-coverage diagnostics on the v4mix training series found
that step 20k retains more multimodal coverage of variant modes (arm
switch, waypoint direction, grasp point) than later steps 35k+, which
progressively collapse toward the modal action. For downstream residual
adapter / noise-steering experiments that rely on the base policy having
non-trivial support over variant modes, 20k is a better starting point
than converged 35k+.

For strict in-distribution BC accuracy on canonical demos, later
checkpoints (35k / 50k) may be preferred.

## Closed-loop sanity (small smoke)

On `beat_block_hammer_D_wpN` val split (arx-x5, SAPIEN sim), 3 seeds:
2 successes out of 2 solvable seeds (1 seed marked expert_fail by the
planner and skipped). Full closed-loop success rate table has not been
computed yet.

## License

Model weights are derived from Physical Intelligence's pi0.5 base and
inherit its Apache-2.0 license. Training data (RoboTwin) has its own
license — see the [RoboTwin repo](https://github.com/RoboTwin-Platform/RoboTwin).