Adaptive Control of Autonomous Vehicles in Dynamic Environments

Abstract: Autonomous vehicles (AVs) have emerged as a significant advancement in modern transportation systems, promising safer and more efficient mobility. However, their large-scale deployment remains challenging due to the inherent risks of operating in non-stationary and unpredictable environments. These challenges stem from the vehicles' limited ability to generalize across unfamiliar dynamics and adapt to sudden environmental changes, such as varying road conditions, traffic behaviors, and weather patterns. The inability to effectively handle such non-stationary conditions poses risks to both safety and reliability.

In this paper, we propose a framework that continuously monitors the vehicle’s dynamics to detect non-stationarities and adapt control policies accordingly. Specifically, we formulate the problem as a Markov Decision Process (MDP), modeling the stochastic interactions between the vehicle and its environment. To enable continuous adaptation, we leverage Deep Reinforcement Learning (DRL) techniques, allowing the framework to capture non-stationary conditions that impact state transitions and dynamically adjust the control policy in response to these changes. This adaptive control approach addresses the critical need for AVs to react in real-time to evolving circumstances and operate safely under uncertainty, ensuring robust decision-making and reliable performance. The effectiveness of the proposed framework will be demonstrated through simulations in the CARLA autonomous driving environment, showcasing the vehicle’s ability to maintain stable and efficient control across a range of dynamic scenarios.