E-MPC Mixture Optimization: A Learning-Enhanced Framework for Autonomous Vehicle Trajectory Planning

Basil Hasan Khwaja

Improves classical Model Predictive Control by replacing uniform motion-primitive sampling with an adaptive mixture distribution balancing exploration and exploitation. The method combines historical trajectory performance with uniform sampling through a tunable mixing coefficient β, allowing the planner to dynamically adapt to changing terrain conditions. β is optimized in real time via a confidence-aware heuristic with stabilized online updates and a NeuralUCB contextual bandit using Bayesian uncertainty estimation and autoregressive temporal context. Experiments in simulated autonomous driving environments demonstrate reduced cumulative trajectory cost and improved sampling efficiency over uniform MPC and heuristic-only baselines, while maintaining computational feasibility for edge-assisted deployment.