Deception Game: Closing the safety-learning loop in interactive robot autonomy

Haimin Hu¹, Zixu Zhang¹, Kensuke Nakamura, Andrea Bajcsy, Jaime F. Fisac

¹equal contribution

Published as a conference paper at CoRL'2023.

Table of Contents

1. About The Project
2. Installation
3. Training
4. Closed-loop Simulation
5. License
6. Contact
7. Citation

About The Project

This repository implements a general RL-based framework for approximate HJI Reachability analysis in joint physical-belief spaces. The resulting control policies explicitly account for a robot's ability to learn and adapt at runtime. The repository is primarily developed and maintained by Haimin Hu and Zixu Zhang.

Click to watch our spotlight video:

Installation

This repository relies on ISAACS. Please follow the instructions there to set up the environment.

Training

Please follow these steps to train the control and disturbance/adversary policies.

• Pretrain a control policy

  python script/bgame_intent_pretrain_ctrl.py

• Pretrain a disturbance policy

  python script/bgame_intent_pretrain_dstb.py

• Joint control-disturbance training

  python script/bgame_intent_isaacs.py

To train the baseline policies, replace bgame with robust and repeat the above steps.

Closed-loop Simulation

We provide a Notebook for testing the trained policies in closed-loop simulations and comparing to baselines.

License

Distributed under the MIT License. See LICENSE for more information.

Contact

Haimin Hu - @HaiminHu - [email protected]

Citation

If you found this repository helpful, please consider citing our paper.

@inproceedings{hu2023deception,
  title={Deception game: Closing the safety-learning loop in interactive robot autonomy},
  author={Hu, Haimin and Zhang, Zixu and Nakamura, Kensuke and Bajcsy, Andrea and Fisac, Jaime Fern{\'a}ndez},
  booktitle={7th Annual Conference on Robot Learning},
  year={2023}
}