---
dataset_info:
  features:
  - name: uid
    dtype: string
  - name: body
    sequence:
      sequence: int64
  - name: connections
    sequence:
      sequence: int64
  - name: reward
    dtype: float64
  - name: env_name
    dtype: string
  - name: generated_by
    dtype: string
  - name: policy_blob
    dtype: binary
  splits:
  - name: train
    num_bytes: 203871816
    num_examples: 2553
  download_size: 201084330
  dataset_size: 203871816
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
license: cc-by-nc-4.0
task_categories:
- robotics
tags:
- robotics
- soft-robotics
- voxel-robots
- reinforcement learning
size_categories:
- 1K<n<10K
---

Evolution Gym is a large-scale benchmark for co-optimizing the design and control of soft robots. It provides a lightweight soft-body simulator wrapped with a gym-like interface for developing learning algorithms. EvoGym also includes a suite of 32 locomotion and manipulation tasks, detailed on our [website](https://evolutiongym.github.io/all-tasks). Task suite evaluations are described in our [NeurIPS 2021 paper](https://arxiv.org/pdf/2201.09863).

<img src="https://github.com/EvolutionGym/evogym/raw/main/images/teaser-low-res.gif" alt="teaser" style="width: 50%; display: block; margin: auto;" />

In this dataset, we open-source 2.5k+ annotated robot structures and policies from the EvoGym paper. The fields of each robot in the dataset are as follows:

- `uid` *(str)*: Unique identifier for the robot [[1]](#note1)  
- `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 [[2]](#note2)  
- `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


<span id="note1">[1]</span> This dataset is a subset of [EvoGym/robots](https://huggingface.co/datasets/EvoGym/robots)  
<span id="note2">[2]</span> Rewards may not exactly match those in [EvoGym/robots](https://huggingface.co/datasets/EvoGym/robots), due to changes in the library, system architecture, etc.


If you find this dataset helpful to your research, please cite 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}
}
```