Hyundai Uiwang — Diffusion Policy (FlowMatch, DINOv3-base) — RIGHT
LeRobot Diffusion Policy, FlowMatch scheduler (num_inference_steps=1), vision encoder
facebook/dinov3-vitb16-pretrain-lvd1689m (ViT-B/16, pretrained+frozen).
Long action chunk: horizon=64, n_action_steps=30. Robot: HDRB + INSPIRE-RH56 (right).
- Model inputs: front_rgb + wrist_rgb images + observation.state(26)
- Output: action(26) = arm 6 + hand 20.
- Trained on the Hyundai Uiwang right-arm dataset (118 ep / 77,490 frames, 30 Hz).
Load-time requirements
- DINOv3 gated access — loading builds the backbone via
AutoModel.from_pretrained("facebook/dinov3-vitb16-pretrain-lvd1689m") (accept its license first).
- lerobot fork with: DINOv3 register-token support.
