Probabilistic Adaptation for Robust Decision-Making of Autonomous Vehicles in Non-Stationary Environments

Autonomous vehicles operate in highly dynamic and uncertain environments where real-world conditions often deviate significantly from those encountered during the training of control policies. Robustness and autonomy in such settings demand policies that can identify the changes and adapt accordingly in real time. This paper presents a framework that integrates pre-trained deep reinforcement learning policies with probabilistic reasoning to enable robust and real-time decision-making in non-stationary environments. The pre-training phase consists of a set of policies trained for specific environments and conditions, which are integrated and deployed in real time. The integration is through probabilistic reasoning, which quantifies the alignment of observed data with each pre-trained model and enables dynamic adaption to ensure maximizing performance while accounting for risk. Two metrics are introduced to assess the robustness and consistency of the performed policy given the probabilistic knowledge, enabling multi-level evaluation of the policy without knowledge of the true environment. Experiments conducted in CARLA demonstrate the effectiveness of the proposed framework in adaptive cruise control across diverse driving scenarios.