README.md

---
license: apache-2.0
library_name: lerobot
tags:
  - robotics
  - imitation-learning
  - aloha
  - diffusion-policy
  - lerobot
  - baseline
datasets:
  - lerobot/aloha_sim_insertion_human_image
pipeline_tag: robotics
---

# Diffusion Policy for ALOHA Insertion Task (Baseline)

⚠️ **Note: This model underperforms ACT on this task. Published for comparison purposes.**

A Diffusion Policy model trained on the ALOHA simulation Insertion task. This model is published as a **baseline comparison** to demonstrate that ACT outperforms Diffusion Policy on ALOHA bimanual tasks.

## Key Finding

| Model | Steps | Success Rate | Task Difficulty |
|-------|-------|--------------|-----------------|
| **ACT** | 200K | **15%** | Hard |
| Diffusion Policy | 200K | 10% | Hard |

**Conclusion: ACT is the recommended approach for ALOHA tasks.**

## Model Description

| Property | Value |
|----------|-------|
| Architecture | Diffusion Policy |
| Parameters | ~100M |
| Task | ALOHA Insertion-v0 |
| Training Steps | 200,000 |
| Batch Size | 32 |
| Success Rate | 0-10% |

## Training Data

- **Dataset**: [lerobot/aloha_sim_insertion_human_image](https://huggingface.co/datasets/lerobot/aloha_sim_insertion_human_image)
- **Episodes**: 50 human demonstrations
- **Frames**: 20,000

## Task Description

The Insertion task requires a bimanual robot to:
1. Pick up a socket with the left arm
2. Pick up a peg with the right arm
3. Insert the peg into the socket in mid-air

⚠️ **This is a difficult task** requiring precise bimanual coordination.

## Demo Video

<video controls src="eval_episode_3.mp4" title="Insertion Diffusion Policy Demo"></video>

## Training Environment

- **GPU**: RTX A6000
- **Framework**: LeRobot 0.4.3
- **Training Time**: Around 12 hours

## Usage

### Installation
```bash
pip install lerobot gym-aloha
```

### Training
```bash
lerobot-train \
    --policy.type=diffusion \
    --dataset.repo_id=lerobot/aloha_sim_insertion_human_image \
    --env.type=aloha \
    --env.task=AlohaInsertion-v0 \
    --batch_size=32 \
    --steps=200000 \
    --eval.n_episodes=10 \
    --eval_freq=20000 \
    --save_freq=20000 \
    --output_dir=./outputs/dp_aloha_insertion \
    --wandb.enable=false \
    --policy.push_to_hub=false
```

### Evaluation
```bash
lerobot-eval \
    --policy.path=LeTau/diffusion_aloha_insertion \
    --env.type=aloha \
    --env.task=AlohaInsertion-v0 \
    --eval.batch_size=1 \
    --eval.n_episodes=20
```

## Results

| Evaluation | Episodes | Success Rate | Avg Sum Reward |
|------------|----------|--------------|----------------|
| Training (200K) | 10 | 10% | 25.0 |
| Independent | 20 | 0% | 17.4 |

**Expected success rate: 0-10%**

## Detailed Evaluation Results (Independent)
```
Sum Rewards: [0.0, 0.0, 37.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
              0.0, 0.0, 0.0, 311.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]

Successes: 0/20 episodes
```


## Comparison: ACT vs Diffusion Policy on ALOHA Tasks

| Task | ACT | Diffusion Policy |
|------|-----|------------------|
| TransferCube (Easy) | **42%** | 10% |
| Insertion (Hard) | **15%** | 0% |

**ACT consistently outperforms Diffusion Policy on ALOHA bimanual tasks.**

## Why Does Diffusion Policy Underperform?

1. **ACT is designed for ALOHA**: ACT was specifically created for bimanual manipulation tasks
2. **Data efficiency**: Diffusion Policy may need more demonstrations to learn effectively
3. **Task characteristics**: ALOHA tasks require precise, deterministic actions rather than multi-modal action distributions

## Recommendation

For ALOHA bimanual tasks, use **ACT** instead:
- [LeTau/act_aloha_transfer_cube](https://huggingface.co/LeTau/act_aloha_transfer_cube) - 42% success rate
- [LeTau/act_aloha_insertion](https://huggingface.co/LeTau/act_aloha_insertion) - 15% success rate

## Citation
```bibtex
@article{zhao2023learning,
  title={Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware},
  author={Zhao, Tony Z and Kumar, Vikash and Levine, Sergey and Finn, Chelsea},
  journal={arXiv preprint arXiv:2304.13705},
  year={2023}
}

@article{chi2023diffusion,
  title={Diffusion Policy: Visuomotor Policy Learning via Action Diffusion},
  author={Chi, Cheng and Feng, Siyuan and Du, Yilun and Xu, Zhenjia and Cousineau, Eric and Burchfiel, Benjamin and Song, Shuran},
  journal={arXiv preprint arXiv:2303.04137},
  year={2023}
}
```

## Acknowledgments

- [LeRobot](https://github.com/huggingface/lerobot) framework by HuggingFace
- [ALOHA](https://tonyzhaozh.github.io/aloha/) project by Stanford
- [Diffusion Policy](https://diffusion-policy.cs.columbia.edu/) project by Columbia