dp-clean-delta-ee
DiffusionPolicy trained on clean bimanual robot manipulation data. Action space: delta end-effector poses (Δright_wrist(7) + Δleft_wrist(7) + grippers(2) = 16-dim). 142 episodes, 45116 frames @ 20fps. ResNet18 backbone, 2 cameras (cam_head + cam_wrist), 279M params, 100k training steps.
Training Details
| Parameter | Value |
|---|---|
| Policy | DiffusionPolicy (UNet) |
| Vision backbone | ResNet18 |
| Cameras | cam_head + cam_wrist (640×480) |
| Obs steps | 2 |
| Action horizon | 16 |
| Action steps | 8 |
| Batch size | 64 |
| Steps | 100,000 |
| Optimizer | Adam (lr=1e-4, betas=[0.95,0.999]) |
| Scheduler | Cosine (500 warmup steps) |
| Noise scheduler | DDIM (100 train / 10 inference steps) |
| Normalization | MEAN_STD (visual), MIN_MAX (state/action) |
Dataset
- Split: clean_train (142 episodes, 45,116 frames @ 20 Hz)
- Robot: Bimanual SO-101 (2× 7-DoF arms)
- Task: Clean object manipulation
Usage
from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
policy = DiffusionPolicy.from_pretrained("atharva-manav/dp-clean-delta-ee")