---
license: cc-by-nc-4.0
tags: [robotics, vla, pi05, openpi, manipulation]
---

# hist3

Three frames at 0.5 s spacing plus a relative-motion state (recent pose deltas rather than an absolute pose).

| | |
|---|---|
| training config | `pi05_frame_hist3` |
| checkpoint step | 7,279 of 7,280 |
| samples seen | 232,928 |
| batch size | 32 |
| peak LR | 0.0005, cosine over 7,280 steps |
| conditioning | 3 frame(s), state `relative_history` |
| action space | relative EE, `world_rotvec`, horizon 50 |
| base weights | `pi05_base` (PyTorch), LoRA r16 `all_linear` |
| frame yaw aug | none |
| normalizer key | `mix/tuned_h50_relative_history3` |

## Contents

- `model.safetensors` — the weights
- `norm_stats.json` — the quantile normalizer this checkpoint was trained against. Not
  interchangeable between variants: the absolute-pose and relative-motion state
  representations have different scales, and using the wrong one silently rescales output.
- `train_config.json` — the resolved training config
- `mixture.json` — dataset sampling weights

## Loading

```python
from experiments.infer import VariantPolicy
policy = VariantPolicy.load("hist3")
actions = policy.infer(images, poses, prompt)   # absolute EE targets, [horizon, 10]
```

## Provenance

Trained from `pi05_base` (Physical Intelligence openpi) on cong-lab manipulation data plus a
DROID subset. Inherits the licence terms of both.

## Caveats

Trained on a 12-hour single-GPU budget, which is roughly 0.05 of one pass
over the 4,245,385-sample mixture. These are ablation checkpoints for a specific
comparison, not a converged production policy.