---
license: apache-2.0
library_name: lerobot
pipeline_tag: robotics
tags:
- robotics
- so-101
- diffusion-policy
- multi-task-dit
- dinov3
- towel-folding
base_model: facebook/dinov3-vitb16-pretrain-lvd1689m
datasets:
- larsvandorp/magic_soup
---

# folding_dit — Multi-Task DiT (DINOv3-B) for towel folding

Diffusion-Transformer policy for autonomous towel folding on a 6-DoF **SO-101** follower arm with a single wrist camera. This is the model used at the competition (ETH "Robot Learning: From Fundamentals to Foundation Models", Project 5 — Diffusion Policy).

The repo **root holds the step-28000 checkpoint** (the deployed one), so `from_pretrained("larsvandorp/folding_dit")` loads it directly.

## Architecture

| Component | Spec |
|---|---|
| Vision encoder | **DINOv3 ViT-B/16** (~86M, fine-tuned, lr × 0.1) |
| Text encoder | CLIP ViT-B/16 text tower (frozen, learnable projection) |
| Noise predictor | 6-layer DiT, 512 hidden, 8 heads, AdaLN-Zero, RoPE |
| Objective | Diffusion, **DDIM** scheduler — 100 train timesteps, 10-step inference |
| Horizon / action steps | 32 / 24 (≈1.0 s / 0.8 s at 30 Hz) |
| Augmentation | resize-only (no crop) + RandomGrayscale (≈50% of samples) + color jitter, no rotation |

Requires the `multi_task_dit` policy + DINOv3 `AutoModel` loading, which live in the fork
[`LarsvanDorp/lerobot@dinov3`](https://github.com/LarsvanDorp/lerobot/tree/dinov3) (not yet upstream).

## Run it

```bash
uv venv --python 3.12 .venv
GIT_LFS_SKIP_SMUDGE=1 uv pip install --python .venv/bin/python \
  "lerobot[multi_task_dit] @ git+https://github.com/LarsvanDorp/lerobot.git@dinov3"
.venv/bin/hf download facebook/dinov3-vitb16-pretrain-lvd1689m   # gated — accept the license first
.venv/bin/hf download openai/clip-vit-base-patch16

.venv/bin/lerobot-rollout \
  --strategy.type=base \
  --robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=my_follower \
  --robot.cameras="{wrist: {type: opencv, index_or_path: 0, width: 800, height: 600, fps: 30, fourcc: MJPG}}" \
  --policy.path=larsvandorp/folding_dit \
  --policy.device=cuda --inference.type=sync \
  --task="fold the towel" --duration=60
```

Feed **color** RGB (the "rgray" model trained on a color+grayscale mix). We always run **without** `--interpolation_multiplier`. Runs on Mac MPS too (drop `fourcc: MJPG`, set `--policy.device=mps`).

## Training data

[`larsvandorp/magic_soup`](https://huggingface.co/datasets/larsvandorp/magic_soup) — ~430 SO-101 towel-folding episodes, deliberately broad (cloths, rotations, locations), grasp next-to-corner, return-to-start after first fold.