LHADRO: A Robust Control Framework for Autonomous Vehicles Under Cyber-Physical Attacks

Deep reinforcement learning has demonstrated remarkable performance in autonomous vehicle control. However, the increasing threat of cyber-physical attacks, which can alter sensor information or vehicle dynamics, poses significant challenges to the robustness of these control policies. To address this, we propose LHADRO (Lambda-History Aware Diversity Robust Oracle), a novel framework that models robust control as a two-player game between control policies and cyber-physical attacks. The key contributions of LHADRO are: (1) A lambda-history aware mechanism that balances past and present meta-policies to enhance training efficiency and mitigates meta-policy thrashing, and (2) A joint diversity introduction mechanism that improves robust control performance by increasing population disparity through a regularization term in parameter updates. We validate the proposed method in MetaDrive-based environments. Experiment results verify that the LHADRO framework improves the robust control performance, and the effectiveness of some critical factors is investigated and discussed.