Autonomous vehicle control based on diffusion policy learning via Gaussian Process Regression in the presence of uncertainties

Enhancing the autonomy of autonomous vehicles (AVs) is an active area of research, as AVs exhibit the potential to improve traffic efficiency by mitigating stop-and-go behaviour in traffic systems. In real-world settings, AVs tend to operate in traffic environments characterised by non-stationary dynamics, where traffic conditions can be naturally changed by stochastic human driving behaviour. However, most existing AV control methods are developed under stationary traffic dynamics and do not explicitly account for environmental shifts, which limits their effectiveness in non-stationary traffic. This paper addresses the problem of AV control under non-stationary traffic conditions by proposing a novel policy approximation algorithm that explicitly accounts for environmental shifts. The proposed approach employs diffusion models within a behaviour cloning framework to approximate the control policy from expert demonstrations. Additionally, Gaussian Process Regression acts as an uncertainty-detection mechanism for the control policy. The proposed method improves control performance in the presence of non-stationary traffic dynamics. The effectiveness of the proposed approach is evaluated in a mixed-autonomy traffic system involving both autonomous and human-driven vehicles. Simulation results demonstrate that the proposed method achieves an approximately 30% increase in overall vehicle speed under non-stationary traffic dynamics compared to baseline methods.