---
tags:
- robotics
- vla
- unlearning
- pi0.5
- libero
license: apache-2.0
---

# Behavior Uncloning — KL (step 20)

VLA unlearning checkpoint: pi0.5 model with **KL** unlearning applied.

## Results

| Metric | Value |
|--------|-------|
| Method | KL |
| Training Steps | 20 |
| Forget Task | "turn on the stove" (LIBERO-Goal T6) |
| Forget SR | 40% (baseline: 100%) |
| Retain SR | 95.6% (baseline: 97.8%) |
| HM | 0.81 |

## Usage

```bash
# Serve with openpi
uv run scripts/serve_policy.py --env LIBERO policy:checkpoint \
    --policy.config pi05_libero --policy.dir <path_to_checkpoint>
```

## Method

**KL Minimization**: L = -L_forget + γ·L_retain + γ·||L_cur - L_orig||². Anchors retain behavior to original model.

Base model: [pi0.5 LIBERO](https://github.com/Physical-Intelligence/openpi)

See full report: [experiment_report.md](https://github.com/haohww/behavior-uncloning/blob/main/docs/experiments/vla/experiment_report.md)