Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Behavior Uncloning - Baseline DiffusionPolicy
|
| 2 |
+
|
| 3 |
+
Baseline DiffusionPolicy trained on all LIBERO-10 episodes (379 episodes, ~101K frames).
|
| 4 |
+
|
| 5 |
+
## Checkpoints
|
| 6 |
+
- `step_025000/` - 25K steps, 76% overall success rate
|
| 7 |
+
- `step_050000/` - 50K steps, 76% overall success rate (final)
|
| 8 |
+
|
| 9 |
+
## Training Config
|
| 10 |
+
- Model: DiffusionPolicy (267M params)
|
| 11 |
+
- Dataset: HuggingFaceVLA/libero (LIBERO-10 subset)
|
| 12 |
+
- Effective batch size: 64 x 8 GPUs = 512
|
| 13 |
+
- Training time: ~20 hours on 8xA100-80GB
|
| 14 |
+
- Framework: LeRobot v0.4.4
|
| 15 |
+
|
| 16 |
+
## Per-Task Success Rates (step 50K)
|
| 17 |
+
|
| 18 |
+
| T0 | T1 | T2 | T3 | T4 | T5 | T6 | T7 | T8 | T9 |
|
| 19 |
+
|----|----|----|----|----|----|----|----|----|-----|
|
| 20 |
+
| 10%| 60%| 80%|100%| 80%|100%| 60%|100%| 70%|100%|
|
| 21 |
+
|
| 22 |
+
## Usage
|
| 23 |
+
```python
|
| 24 |
+
from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
|
| 25 |
+
policy = DiffusionPolicy.from_pretrained("haohw/behavior-uncloning-baseline", subfolder="step_050000")
|
| 26 |
+
```
|
| 27 |
+
|
| 28 |
+
## Project
|
| 29 |
+
Part of the [behavior-uncloning](https://github.com/haohww/behavior-uncloning) project — machine unlearning for robot manipulation policies.
|