| tags: | |
| - robotics | |
| - vla | |
| - unlearning | |
| - pi0.5 | |
| - libero | |
| license: apache-2.0 | |
| # Behavior Uncloning — GD (step 20) | |
| VLA unlearning checkpoint: pi0.5 model with **GD** unlearning applied. | |
| ## Results | |
| | Metric | Value | | |
| |--------|-------| | |
| | Method | GD | | |
| | Training Steps | 20 | | |
| | Forget Task | "turn on the stove" (LIBERO-Goal T6) | | |
| | Forget SR | 0% (baseline: 100%) | | |
| | Retain SR | 95.6% (baseline: 97.8%) | | |
| | HM | 0.98 | | |
| ## 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 | |
| **Gradient Difference**: L = -L_forget + L_retain. Adds retain regularization to prevent catastrophic forgetting. | |
| 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) | |