YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Behavior Uncloning - Retrain Oracle DiffusionPolicy
Retrain oracle (upper bound) trained on D_retain only — LIBERO-10 minus T8 episodes (344 episodes, ~93K frames).
Checkpoints
step_027500/- 27.5K steps, 72% overall success ratestep_040000/- 40K steps, 76% overall success rate (best)
Training Config
- Model: DiffusionPolicy (267M params)
- Dataset: HuggingFaceVLA/libero (LIBERO-10 minus T8)
- Forget task: T8 (35 episodes excluded)
- Effective batch size: 64 x 8 GPUs = 512
- Training time: ~16 hours on 8xA100-80GB
- Framework: LeRobot v0.4.4
Per-Task Success Rates (step 40K)
| T0 | T1 | T2 | T3 | T4 | T5 | T6 | T7 | T8 | T9 |
|---|---|---|---|---|---|---|---|---|---|
| 10% | 90% | 90% | 0% | 100% | 100% | 100% | 100% | 90% | 80% |
Key finding: T8 achieves ~90% SR despite being excluded from training, indicating the model generalizes to T8 from other tasks. T3 drops to 0% (collateral damage).
Usage
from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
policy = DiffusionPolicy.from_pretrained("haohw/behavior-uncloning-retrain-oracle", subfolder="step_040000")
Project
Part of the behavior-uncloning project — machine unlearning for robot manipulation policies.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support