Code for working with the EvoGym paper's data. All datasets are hosted on huggingface.

• EvoGym/robots: 90k+ annotated robot structures
• EvoGym/robot-with-policies: 2.5k+ annotated robot structures and policies

The package installed by this repo is necessary to deserialize robot policies. Python versions 3.8 through 3.10 are supported.

conda create -n "egdatasets" python=3.8
conda activate egdatasets
pip install -e .

The EvoGym/robots dataset contains 90k+ annotated robot structures. The fields of each robot in the dataset are as follows:

• uid (str): Unique identifier for the robot
• body (int64 np.ndarray): 2D array indicating the voxels that make up the robot
• connections (int64 np.ndarray): 2D array indicating how the robot's voxels are connected. In this dataset, all robots are fully-connected, meaning that all adjacent voxels are connected
• reward (float): reward achieved by the robot's policy
• env_name (str): Name of the EvoGym environment (task) the robot was trained on
• generated_by ("Genetic Algorithm" | "Bayesian Optimization" | "CPPN-NEAT"): Algorithm used to generate the robot

The EvoGym/robots-with-policies dataset contains 2.5k+ annotated robot structures and their policies. This dataset is a subset of the EvoGym/robots dataset, with the addition of serialized policies for each robot. The fields of each robot in the dataset are as follows:

• uid (str): Unique identifier for the robot
• body (int64 np.ndarray): 2D array indicating the voxels that make up the robot
• connections (int64 np.ndarray): 2D array indicating how the robot's voxels are connected. In this dataset, all robots are fully-connected, meaning that all adjacent voxels are connected.
• reward (float): reward achieved by the robot's policy [1]
• env_name (str): Name of the EvoGym environment (task) the robot was trained on
• generated_by ("Genetic Algorithm" | "Bayesian Optimization" | "CPPN-NEAT"): Algorithm used to generate the robot
• policy_blob (binary): Serialized policy for the robot

[1] Rewards may not exactly match those in EvoGym/robots, due to changes in the library, system architecture, etc.

If you find these datasets useful, please consider citing our paper:

@article{bhatia2021evolution,
  title={Evolution gym: A large-scale benchmark for evolving soft robots},
  author={Bhatia, Jagdeep and Jackson, Holly and Tian, Yunsheng and Xu, Jie and Matusik, Wojciech},
  journal={Advances in Neural Information Processing Systems},
  volume={34},
  year={2021}
}