Risk Averse Robust Adversarial Reinforcement Learning

Xinlei Pan, Daniel Seita, Yang Gao, John Canny

Risk Averse Robust Adversarial Reinforcement Learning

Xinlei Pan, Daniel Seita, Yang Gao, John Canny

ICRA 2019

Updates:

  • 01/31/2019: The paper has been accepted by ICRA 2019.
  • 09/15/2018: The paper has been submitted to ICRA 2019.
  • 09/21/2018: Supplementary video has been submitted

Training Environment:

  • py_TORCS: py_TORCS is an autonomous driving simulation environment developed on top of TORCS, it is compatible with OpenAI gym. See here for more information: https://github.com/xinleipan/py_TORCS

Training Results:

Testing all models without attacks or perturbations. The reward is divided into distance related reward (left subplot), progress related reward (middle subplot). We also present results for catastrophe reward per episode (right subplot). The blue vertical line indicates the beginning of adding perturbations during training.

Testing all models with random attacks. The three subplots follow the same convention as in the first figure.

Testing all models with adversarial attack. The three subplots follow the same convention as in the first figure.

combined.mp4