Haimin Hu1, Zixu Zhang1, Kensuke Nakamura, Andrea Bajcsy, Jaime F. Fisac
1equal contribution
Published as a conference paper at CoRL'2023.
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:
This repository relies on ISAACS. Please follow the instructions there to set up the environment.
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.
We provide a Notebook for testing the trained policies in closed-loop simulations and comparing to baselines.
Distributed under the MIT License. See LICENSE for more information.
Haimin Hu - @HaiminHu - haiminh@princeton.edu
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}
}