CrowdPlay: Crowdsourcing Human Demonstrations for Offline Learning

Matthias Gerstgrasser, Rakshit Trivedi, David Parkes

• Paper
• Code
• Dataset
• Documentation

CrowdPlay is a platform for crowdsourcing human demonstration trajectories at scale in RL environments. It interfaces with standard RL simulators that implement an OpenAI Gym-style API consisting of a reset and a step function. CrowdPlay can be used to generate large-scale datasets of human demonstrations with minimal effort, as detailed in our publication CrowdPlay: Crowdsourcing Human Demonstrations for Offline Learning Matthias Gerstgrasser, Rakshit Trivedi, David C. Parkes

If you use our platform or dataset in your publication, we ask that you cite it using this bibtex entry:

@inproceedings{
gerstgrasser2022crowdplay,
title={CrowdPlay: Crowdsourcing Human Demonstrations for Offline Learning},
author={Matthias Gerstgrasser and Rakshit Trivedi and David C. Parkes},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=qyTBxTztIpQ}
}

Quick Start

1. Prerequisites

Install and start Docker. Clone this repository.

2. Running the app

Run docker-compose -f docker-compose.dev.yaml up to start the project. If you have previously run the project and would like to update all underlying packages, run the following:

docker-compose -f docker-compose.dev.yaml down && rm -r data && docker-compose -f docker-compose.dev.yaml build && docker-compose -f docker-compose.dev.yaml up -d

Afterwards, go to http://127.0.0.1:9000 to start the app. Also check out the documentation.

Acknowledgements

We thank Francisco Ramos for his help in implementing the CrowdPlay software, especially on the frontend. We would like to thank anonymous reviewers at ICLR 2022 for their constructive feedback and discussions. This research is funded in part by Defense Advanced Research Projects Agency under Cooperative Agreement HR00111920029. The content of the information does not necessarily reflect the position or the policy of the Government, and no official endorsement should be inferred. This is approved for public release; distribution is unlimited.