KL Minimization (KL) Step 40 — Object-Level Unlearning
Best checkpoint for KL on LIBERO-Object (forget T7 "butter").
Results
| Metric | Value |
|---|---|
| Total SR | 30% |
| Forget SR (butter) | 0% |
| Retain SR (9 objects) | 33.3% |
| HM | 0.50 |
Usage
# With openpi
uv run scripts/serve_policy.py --env LIBERO policy:checkpoint \
--policy.config pi05_libero --policy.dir <path_to_checkpoint>
Details
- Base model: pi0.5 (3.35B params, flow matching)
- Dataset: LIBERO-Object (10 pick-and-place tasks)
- Forget target: T7 "pick up the butter and place it in the basket"
- Training: 40 steps, BS=32, lr=1e-5
- Report: Object-Level Experiment Report