Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-4.0
|
| 3 |
+
tags:
|
| 4 |
+
- robotics
|
| 5 |
+
- unlearning
|
| 6 |
+
- pi0.5
|
| 7 |
+
- libero
|
| 8 |
+
- object-level
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# KL Minimization (KL) Step 40 — Object-Level Unlearning
|
| 12 |
+
|
| 13 |
+
Best checkpoint for KL on LIBERO-Object (forget T7 "butter").
|
| 14 |
+
|
| 15 |
+
## Results
|
| 16 |
+
|
| 17 |
+
| Metric | Value |
|
| 18 |
+
|--------|-------|
|
| 19 |
+
| Total SR | 30% |
|
| 20 |
+
| Forget SR (butter) | 0% |
|
| 21 |
+
| Retain SR (9 objects) | 33.3% |
|
| 22 |
+
| HM | 0.50 |
|
| 23 |
+
|
| 24 |
+
## Usage
|
| 25 |
+
|
| 26 |
+
```python
|
| 27 |
+
# With openpi
|
| 28 |
+
uv run scripts/serve_policy.py --env LIBERO policy:checkpoint \
|
| 29 |
+
--policy.config pi05_libero --policy.dir <path_to_checkpoint>
|
| 30 |
+
```
|
| 31 |
+
|
| 32 |
+
## Details
|
| 33 |
+
|
| 34 |
+
- Base model: pi0.5 (3.35B params, flow matching)
|
| 35 |
+
- Dataset: LIBERO-Object (10 pick-and-place tasks)
|
| 36 |
+
- Forget target: T7 "pick up the butter and place it in the basket"
|
| 37 |
+
- Training: 40 steps, BS=32, lr=1e-5
|
| 38 |
+
- Report: [Object-Level Experiment Report](https://github.com/haohww/behavior-uncloning/blob/main/docs/experiments/vla/object_level_report.md)
|