Reinforcement Learning Journal, vol. 4, 2024, pp. 2017–2040.
Presented at the Reinforcement Learning Conference (RLC), Amherst Massachusetts, August 9–12, 2024.
Reinforcement learning (RL) holds promise for the next generation of autonomous vehicles, but it lacks formal robustness guarantees against adversarial attacks in the observation space for safety-critical tasks. In particular, for obstacle avoidance tasks, attacks on the observation space can significantly alter vehicle behavior, as demonstrated in this paper. Traditional approaches to enhance the robustness of RL-based control policies, such as training under adversarial conditions or employing worst-case scenario planning, are limited by their policy's parameterization and cannot address the challenges posed by topological obstructions in the presence of noise. We introduce a new hybrid RL algorithm featuring hysteresis-based switching to guarantee robustness against these attacks for vehicles operating in environments with multiple obstacles. This hysteresis-based RL algorithm for coping with multiple obstacles, referred to as MultiHyRL, addresses the 2D bird's-eye view obstacle avoidance problem, featuring a complex observation space that combines local (images) and global (vectors) observations. Numerical results highlight its robustness to adversarial attacks in various challenging obstacle avoidance settings where Proximal Policy Optimization (PPO), a traditional RL method, fails.
Jan de Priester, Zachary Bell, Prashant Ganesh, and Ricardo Sanfelice. "MultiHyRL: Robust Hybrid RL for Obstacle Avoidance against Adversarial Attacks on the Observation Space." Reinforcement Learning Journal, vol. 4, 2024, pp. 2017–2040.
BibTeX:@article{priester2024multihyrl,
title={{MultiHyRL}: {R}obust Hybrid {RL} for Obstacle Avoidance against Adversarial Attacks on the Observation Space},
author={Priester, Jan de and Bell, Zachary and Ganesh, Prashant and Sanfelice, Ricardo},
journal={Reinforcement Learning Journal},
volume={4},
pages={2017--2040},
year={2024}
}