Datasets:
license: mit
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
🛠️ 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
Run run_task.sh to run task:
bash run_task.sh ${task_name} ${gpu_id}
The default task running configurations can be found at config/${task_name}.yml, if you want to change some task setting, just modify the specific configuration file.
RoboTwin will run expert check first to make sure only successful data will be collect (by running seeds one by one first), this will not take too long.
The Data will be collect to data under the root directory, in the form of .pkl usually.
2. Task Config
Data collection configurations are located in the config folder, corresponding to each task. Here is an explanation of the important parameters:
- render_freq: Set to 0 means no rendering. If you wish to see the rendering, it can be set to 10.
- collect_data: Data collection will only be enabled if set to True.
- 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.
- pcd_crop: Determines whether the obtained point cloud data is cropped to remove elements like tables and walls.
- 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.
- data_type/endpose: The 6D pose of the end effector, which still has some minor issues.
- data_type/qpos: Represents the joint action.
- 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
Then, move to policy/3D-Diffusion-Policy first, and run the following code to train DP3 :
bash train.sh ${task_name} ${expert_data_num} ${gpu_id}
Run the following code to eval DP3 for specific task:
bash eval.sh ${task_name} ${expert_data_num} ${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% |
| Block 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% |
| Block Sweep | DP3 (XYZ) | 49% | 80% | 96% |
| | DP3 (XYZ+RGB) | 70% | 98% | 99% |
| Block 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) | 20.15% | 36.46% | 53.23% |
| | DP3 (XYZ+RGB) | 17.93% | 29.33% | 45.93% |
🪄 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
👍 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.