molierflower / RoboTwin

RoboTwin: Dual-Arm Robot Benchmark with Generative Digital Twins

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

RoboTwin: Dual-Arm Robot Benchmark with Generative Digital Twins

Yao Mu* †, Tianxing Chen* , Zanxin Chen* , Shijia Peng*,
Zeyu Gao, Zhiqian Lan, Yude Zou, Lunkai Lin, Zhiqiang Xie, Ping Luo.

RoboTwin (early version), accepted to ECCV Workshop 2024 (Oral): Webpage | PDF | arXiv

📚 Overview

Expert Data Generation

🛠️ Installation

See INSTALLATION.md for installation instructions. It takes about 20 minutes for installation.

ℹ️ Task Informaction

Coming Soon

🧑🏻‍💻 Usage

1. Task Running and Data Collection

This part will be released soon. We now release 50 demos for each task:

In the project root directory:

mkdir data
cd data

View https://huggingface.co/datasets/YaoMarkMu/robotwin_dataset, download the files and unzip them to data.

2. Task Config

Data collection configurations are located in the config folder, corresponding to each task. Here is an explanation of the important parameters:

  1. render_freq: Set to 0 means no rendering. If you wish to see the rendering, it can be set to 10.
  2. collect_data: Data collection will only be enabled if set to True.
  3. camera_w,h: These are the camera parameters, with a total of 4 cameras - two on the wrist and two positioned as top and front views.
  4. pcd_crop: Determines whether the obtained point cloud data is cropped to remove elements like tables and walls.
  5. pcd_down_sample_num: The point cloud data is downsampled using the FPS (Farthest Point Sampling) method, set it to 0 to keep the raw point cloud data.
  6. data_type/endpose: The 6D pose of the end effector, which still has some minor issues.
  7. data_type/qpos: Represents the joint action.
  8. observer: Decides whether to save a observer-view photo for easy observation.

3. Deploy your policy

See envs/base_task.py, search TODO and you may see the following code, make sure that policy.get_action(obs) will return action sequence (predicted actions).:

actions = model.get_action(obs) # TODO, get actions according to your policy and current obs

You need to modify script/eval_policy.py in the root directory to load your model for evaluation: Search TODO, modify the code to init your policy.

Run the follow command to run your policy in specific task env:

bash script/run_eval_policy.sh ${task_name} ${gpu_id}

4. DP3 as baseline

The DP3 code can be found in policy/3D-Diffusion-Policy.

Process Data for DP3 training after collecting data (In root directory), and input the task name and the amount of data you want your policy to train with:

python script/pkl2zarr_dp3.py ${task_name} ${number_of_episodes}

Then, move to policy/3D-Diffusion-Policy first, and run the following code to train DP3 :

bash train.sh ${task_name} ${expert_data_num} ${seed} ${gpu_id}

Run the following code to eval DP3 for specific task:

bash eval.sh ${task_name} ${expert_data_num} ${checkpoint_num} ${seed} ${gpu_id}

Current leaderboard:

Here's the revised table with the averages listed at the end:

|     Task                       |     Algorithm       |   10 demos   |   20 demos   |   50 demos   |
|--------------------------------|---------------------|--------------|--------------|--------------|
| Apple Cabinet Storage          | DP3 (XYZ)           | 41%          | 59%          | 75%          |
|                                | DP3 (XYZ+RGB)       | 22%          | 41%          | 60%          |
| Block Handover                 | DP3 (XYZ)           | 55%          | 89%          | 70%          |
|                                | DP3 (XYZ+RGB)       | 48%          | 81%          | 94%          |
| Blocks Stack (Easy)            | DP3 (XYZ)           | /            | /            | /            |
|                                | DP3 (XYZ+RGB)       | 0%           | 1%           | 23%          |
| Container Place                | DP3 (XYZ)           | 34%          | 54%          | 68%          |
|                                | DP3 (XYZ+RGB)       | 18%          | 28%          | 54%          |
| Dual Bottles Pick (Easy)       | DP3 (XYZ)           | 10%          | 48%          | 78%          |
|                                | DP3 (XYZ+RGB)       | 9%           | 41%          | 75%          |
| Empty Cup Place                | DP3 (XYZ)           | 3%           | 30%          | 73%          |
|                                | DP3 (XYZ+RGB)       | 7%           | 23%          | 82%          |
| Pick Apple Messy               | DP3 (XYZ)           | 2%           | 2%           | 9%           |
|                                | DP3 (XYZ+RGB)       | 2%           | 3%           | 25%          |
| Shoes Place                    | DP3 (XYZ)           | 2%           | 1%           | 12%          |
|                                | DP3 (XYZ+RGB)       | 0%           | 0%           | 5%           |
| Block Hammer Beat              | DP3 (XYZ)           | 37%          | 45%          | 60%          |
|                                | DP3 (XYZ+RGB)       | 36%          | 41%          | 73%          |
| Blocks Stack (Hard)            | DP3 (XYZ)           | /            | /            | /            |
|                                | DP3 (XYZ+RGB)       | 0%           | 0%           | 3%           |
| Diverse Bottles Pick           | DP3 (XYZ)           | 3%           | 12%          | 38%          |
|                                | DP3 (XYZ+RGB)       | 0%           | 1%           | 7%           |
| Dual Bottles Pick (Hard)       | DP3 (XYZ)           | 13%          | 29%          | 46%          |
|                                | DP3 (XYZ+RGB)       | 11%          | 26%          | 48%          |
| Mug Hanging                    | DP3 (XYZ)           | 1%           | 9%           | 13%          |
|                                | DP3 (XYZ+RGB)       | 1%           | 2%           | 6%           |
| Shoe Place                     | DP3 (XYZ)           | 12%          | 16%          | 54%          |
|                                | DP3 (XYZ+RGB)       | 13%          | 20%          | 35%          |
| Average                        | DP3 (XYZ)           | 17.75%       | 32.83%       | 49.67%       |
|                                | DP3 (XYZ+RGB)       | 11.93%       | 22.00%       | 42.14%       |

🪄 Digital Twin Generation

Deemos Rodin: https://hyperhuman.deemos.com/rodin

⁉️ Common Issues

If you find you fail to quit the running python process with Crtl + C, just try Ctrl + \.

We found Vulkan is not stable in someoff-screen devices, try reconnecting ssh -X ... if you meet any problem.

Other Common Issues can be found in COMMON_ISSUE

⏱️ Future Plans

  1. Task Code Generation Pipeline.
  2. RoboTwin (Final Version) will be released soon.
  3. Real Robot Data collected by teleoperation.
  4. Tasks env (Data Collection).
  5. More baseline code will be integrated into this repository (RICE, ACT, Diffusion Policy).

👍 Citation

If you find our work useful, please consider citing:

RoboTwin: Dual-Arm Robot Benchmark with Generative Digital Twins (early version), accepted to ECCV Workshop 2024 (Oral)

@article{mu2024robotwin,
  title={RoboTwin: Dual-Arm Robot Benchmark with Generative Digital Twins (early version)},
  author={Mu, Yao and Chen, Tianxing and Peng, Shijia and Chen, Zanxin and Gao, Zeyu and Zou, Yude and Lin, Lunkai and Xie, Zhiqiang and Luo, Ping},
  journal={arXiv preprint arXiv:2409.02920},
  year={2024}
}

🏷️ License

This repository is released under the MIT license. See LICENSE for additional details.

About

RoboTwin: Dual-Arm Robot Benchmark with Generative Digital Twins

License:MIT License


Languages

Language:Python 98.7%Language:Shell 1.3%