---
library_name: lerobot
pipeline_tag: robotics
tags: [lerobot, diffusion-policy, flow-matching, dinov3, robotics, manipulation]
---

# Hyundai Uiwang — Diffusion Policy (FlowMatch), RIGHT arm — **DINOv3 backbone**

LeRobot Diffusion Policy trained on the Hyundai Uiwang **right**-arm dataset
(118 episodes / 77,490 frames, 30 Hz), FlowMatch scheduler (`num_inference_steps=1`).
Robot: HDRB + INSPIRE-RH56. Vision encoder: **facebook/dinov3-vitb16-pretrain-lvd1689m**
(ViT-B/16), pretrained + frozen. Inputs: front_rgb + wrist_rgb + state(26) +
gripper_sensor(30) + wrist_ft_sensor(6) -> action(26).

## Load-time requirements (important)
1. **DINOv3 gated access**: loading constructs the backbone via
   `AutoModel.from_pretrained("facebook/dinov3-vitb16-pretrain-lvd1689m")`, which is gated —
   accept the license on that model page with your HF account first.
2. **lerobot with DINOv3 register-token support**: the `Dinov2Backbone` wrapper must strip
   CLS + register tokens (DINOv3 has 4 registers). Use the snu-hyundai fork branch that
   includes this patch, or the loaded patch features will be misaligned.