Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
tags:
|
| 3 |
+
- robotics
|
| 4 |
+
- vla
|
| 5 |
+
- unlearning
|
| 6 |
+
- pi0.5
|
| 7 |
+
- libero
|
| 8 |
+
license: apache-2.0
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# Behavior Uncloning — GA (step 20)
|
| 12 |
+
|
| 13 |
+
VLA unlearning checkpoint: pi0.5 model with **GA** unlearning applied.
|
| 14 |
+
|
| 15 |
+
## Results
|
| 16 |
+
|
| 17 |
+
| Metric | Value |
|
| 18 |
+
|--------|-------|
|
| 19 |
+
| Method | GA |
|
| 20 |
+
| Training Steps | 20 |
|
| 21 |
+
| Forget Task | "turn on the stove" (LIBERO-Goal T6) |
|
| 22 |
+
| Forget SR | 0% (baseline: 100%) |
|
| 23 |
+
| Retain SR | 44.4% (baseline: 97.8%) |
|
| 24 |
+
| HM | 0.61 |
|
| 25 |
+
|
| 26 |
+
## Usage
|
| 27 |
+
|
| 28 |
+
```bash
|
| 29 |
+
# Serve with openpi
|
| 30 |
+
uv run scripts/serve_policy.py --env LIBERO policy:checkpoint \
|
| 31 |
+
--policy.config pi05_libero --policy.dir <path_to_checkpoint>
|
| 32 |
+
```
|
| 33 |
+
|
| 34 |
+
## Method
|
| 35 |
+
|
| 36 |
+
**Gradient Ascent**: L = -L_flow(D_forget). Maximizes flow matching loss on forget data.
|
| 37 |
+
|
| 38 |
+
Base model: [pi0.5 LIBERO](https://github.com/Physical-Intelligence/openpi)
|
| 39 |
+
|
| 40 |
+
See full report: [experiment_report.md](https://github.com/haohww/behavior-uncloning/blob/main/docs/experiments/vla/experiment_report.md)
|