Xingyu Peng^*,1,2, Chen Gao^*,1,3, Liankai Jin^*,1, Annan Li¹, Si Liu^†,1

Beihang University¹, Zhongguancun Academy², National University of Singapore³

Website: https://buaa-colalab.github.io/BiCoord/

Overview, Tasks & LeaderBoard

Please see https://buaa-colalab.github.io/BiCoord/ for details.

Installation

【Step 1】Clone the Repository

git clone https://github.com/buaa-colalab/BiCoord-Bench.git

【Step 2】Install the RoboTwin 2.0 Environment

Please follow RoboTwin 2.0 Document (Usage - Install & Download) to install the RoboTwin 2.0 environment.

【Step 3】Download Additional Assets

We have modified some objects in the original RoboTwin-OD dataset. Please download the additional objects at https://huggingface.co/datasets/GradiusTwinbee/BiCoord/blob/main/objects.zip and extract them to BiCoord-Bench/assets/objects.

【Step 4】For the env installation of specific policies, please refer to RoboTwin 2.0 Usage Guide - RoboTwin 2.0 Offical Document

In our practice, we install the RoboTwin 2.0 Environment first, and then install the policies' envs respectively based on this basic environment.

Data Preparation

We recommend to download official training data at https://huggingface.co/datasets/GradiusTwinbee/BiCoord and put data under BiCoord-Bench/data. Otherwise, you can generate data following the guidance at Collect Data - RoboTwin 2.0 Offical Document, and run correspnding split_stages.py to obtain stage information.

After preparation, BiCoord-Bench/data should look like

data/
├── balance_roller/
│   └── demo_clean/
│       ├── _traj_data/
│       ├── .ipynb_checkpoints/
│       ├── data/					#Trajectory data, including endpose, joint_action and observation
│       ├── instructions/			#Task instruction
│       ├── stages/					#Stage information for each episode
│       ├── video/					#Video
│       ├── scene_info.json
│       ├── seed.txt				#The corresponding random seeds for producing the trajectory data
│       └── split_stages.py			#Produce stage information based on trajectory data
├── build_tower_with_blocks/
├── clean_table/
├── ......

After get data prepared under BiCoord-Bench/data, you need to convert the data into the training format according to the policy you choose. For data conversion of baseline policies, please follow the guidance at RoboTwin 2.0 Usage Guide - RoboTwin 2.0 Offical Document.

Training & Evaluation

Please follow RoboTwin 2.0 Usage Guide - RoboTwin 2.0 Offical Document.

For convenience, we also provide all checkpoints of baseline policies (DP, RDT, OpenVLA-OFT, Pi0 under single&multi task setting) at https://huggingface.co/Oshwiciqwq/BiCoord-checkpoints.

Acknowledgement

We sincerely thank RoboTwin 2.0 for their outstanding contributions to bimanual manipulation simulation, convenient action APIs and open-source release.

We also thank DP, RDT, OpenVLA-OFT and Pi0 for their outstanding and representive contributions in manipulation and VLAs.

License

This benchmark is released under the MIT License.