README.md

---
license: mit
task_categories:
- robotics
- reinforcement-learning
language: []
pretty_name: LeFold Preview
size_categories:
- 10M<n<100M
tags:
- lerobot
- so101
- dual-arm
- manipulation
- folding
- cloth
---

# LeFold Preview

This is a preview dataset for **LeFold**, a dual-arm cloth folding manipulation dataset collected with the **SO-101** dual-arm robot. The dataset is built using the [LeRobot](https://github.com/huggingface/lerobot) v3.0 format and is fully compatible with LeRobot training pipelines.

## Dataset Summary

| Property | Value |
|----------|-------|
| Robot | SO-101 Dual Arm (`so101_dualarm`) |
| Total Episodes | 1,200 |
| Total Frames | 738,533 |
| Number of Tasks | 4 |
| FPS | 30 |
| Action Dim | 12 (6 per arm) |
| State Dim | 12 (6 joint positions per arm) |
| Arm Spacing | 46 cm (center-to-center of two arm bases) |

### Task Categories

| Task Index | Task | Episodes | Frames |
|------------|------|----------|--------|
| 0 | Fold the long pants (长裤折叠) | 300 | 162,084 |
| 1 | Fold the long sleeves (长袖折叠) | 300 | 246,170 |
| 2 | Fold the short pants (短裤折叠) | 300 | 90,854 |
| 3 | Fold the short sleeves (短袖折叠) | 300 | 239,425 |

## Observation Space

### Cameras

| Camera Key | Resolution | Codec | Description |
|------------|-----------|-------|-------------|
| `observation.images.left_wrist` | 480×640 | AV1 (yuv420p) | Left wrist-mounted camera |
| `observation.images.right_wrist` | 480×640 | AV1 (yuv420p) | Right wrist-mounted camera |
| `observation.images.right_front` | 720×1280 | AV1 (yuv420p) | Right front view camera |

### State

Joint positions (float32 × 12):

| Index | Joint Name |
|-------|------------|
| 0 | `left.shoulder_pan.pos` |
| 1 | `left.shoulder_lift.pos` |
| 2 | `left.elbow_flex.pos` |
| 3 | `left.wrist_flex.pos` |
| 4 | `left.wrist_roll.pos` |
| 5 | `left.gripper.pos` |
| 6 | `right.shoulder_pan.pos` |
| 7 | `right.shoulder_lift.pos` |
| 8 | `right.elbow_flex.pos` |
| 9 | `right.wrist_flex.pos` |
| 10 | `right.wrist_roll.pos` |
| 11 | `right.gripper.pos` |

## Action Space

12-dimensional continuous action (same structure as state, 6 DoF per arm).

## Splits

| Split | Episodes |
|-------|----------|
| Train | 0–1200 |

## Usage

```python
from lerobot.datasets import LeRobotDataset

dataset = LeRobotDataset(
    repo_id="cmriat/lefold-preview",
    split="train",
)
```

Or download directly via `huggingface-cli`:

```bash
huggingface-cli download cmriat/lefold-preview --repo-type dataset
```

### Training

First install LeRobot with training dependencies:

```bash
pip install lerobot[training]
```

#### Method 1: Training Config File (Recommended)

Create a config file `train_lefold.yaml`:

```yaml
dataset:
  repo_id: cmriat/lefold-preview
  video_backend: pyav

policy:
  type: pi05                   # PI0.5 — flow-matching vision-language-action policy
  device: cuda
  push_to_hub: false
  pretrained_path: lerobot/pi05_base   # fine-tune from PI05 base checkpoint

  paligemma_variant: gemma_2b
  action_expert_variant: gemma_300m
  dtype: bfloat16

  chunk_size: 30                # action prediction horizon
  n_obs_steps: 1
  max_state_dim: 32
  max_action_dim: 32
  image_resolution: [224, 224]

  use_relative_actions: true    # predict delta joint positions
  relative_exclude_joints: ["gripper"]

  n_action_steps: 10            # execute first 10 predicted actions
  num_inference_steps: 10       # flow-matching inference steps

  input_features:
    observation.state:
      type: STATE
      shape: [12]
    observation.images.left_wrist:
      type: VISUAL
      shape: [3, 224, 224]
    observation.images.right_front:
      type: VISUAL
      shape: [3, 224, 224]
    observation.images.right_wrist:
      type: VISUAL
      shape: [3, 224, 224]

  output_features:
    action:
      type: ACTION
      shape: [12]

batch_size: 16
steps: 60000
save_freq: 5000
log_freq: 100
num_workers: 8

optimizer:
  type: adamw
  lr: 3e-5
  betas: [0.9, 0.95]
  eps: 1e-8
  weight_decay: 0.01
  grad_clip_norm: 1.0

scheduler:
  type: cosine_decay_with_warmup
  num_warmup_steps: 1200
  num_decay_steps: 60000
  peak_lr: 3e-5
  decay_lr: 3e-6

wandb:
  enable: true
  project: lefold-so101
  disable_artifact: true

output_dir: outputs/lefold_pi05
job_name: lefold_pi05
```

Then launch training:

```bash
lerobot-train --config_path train_lefold.yaml
```

#### Method 2: CLI Only

For quick experiments without a config file:

```bash
lerobot-train \
    --dataset.repo_id=cmriat/lefold-preview \
    --policy.type=pi05 \
    --policy.pretrained_path=lerobot/pi05_base \
    --policy.device=cuda \
    --policy.use_relative_actions=true \
    --batch_size=16 \
    --steps=60000 \
    --output_dir=outputs/lefold_pi05 \
    --wandb.enable=true \
    --wandb.project=lefold-so101
```

For other supported policy types (e.g., ACT, Diffusion, multi-task DIT, SmolVLA), replace `--policy.type` above or refer to the [LeRobot documentation](https://github.com/huggingface/lerobot).

> **Note on SO-101 hardware setup:**
> - The two arms of the SO-101 dual-arm robot are mounted with their base centers **46 cm apart**. When deploying a trained policy, ensure your hardware setup matches this spacing for consistent kinematics.
> - The front camera is positioned **between the two arms**, and the camera views are aligned via **RGB image matching** (i.e., adjusting camera pose until the RGB frames from each camera are visually consistent with the dataset).

## License

MIT