UC Berkeley

Model Predictive Control (MPC) has become a fundamental tool for motion planning and control in autonomous driving, valued for its ability to enforce constraints on states and inputs while optimizing a performance objective over a finite prediction horizon. Even as the autonomous vehicle industry increasingly delegates perception, prediction, and high-level planning to learned models from data, MPC remains a critical component of the control and planning stack at major autonomous driving companies, providing interpretable constraint enforcement, safety verification, and fallback planning that complement the primary neural planner. However, as driving scenarios grow in complexity—interacting agents with coupled decisions, surrounding vehicles exhibiting uncertain multi-modal behavior, and long planning horizons that span multiple traffic lights in urban driving—the optimization problem MPC must solve at each control step becomes computationally intractable for real-time deployment. This dissertation develops computationally efficient MPC algorithms that make optimization-based planning practical and scalable across three distinct areas of complexity in autonomous driving.The first part addresses scalable multi-agent interactive planning using MPC. We propose a distributed MPC coordination framework for cooperative multi-vehicle lane changing, in which a designated facilitator vehicle proactively modifies the traffic environment to enable lane changes for the remaining vehicles in a platoon, yielding a formulation that scales with the number of agents without a corresponding growth in per-agent computational cost. We then address competitive and cooperative two-agent interactions through a game-theoretic lens: a value function learned offline from generalized Nash equilibrium solutions is embedded as the terminal cost-to-go in a short-horizon MPC policy, preserving strategic behavior while reducing online computation to a fixed-size optimal control problem that each agent solves independently.The second part addresses scalable stochastic MPC with multi-modal trajectory predictions of surrounding vehicles, where collision-avoidance constraints grow combinatorially with the number of agents and prediction modes, preventing real-time deployment. We introduce a duality-based screening framework that uses an attention-based recurrent neural network to predict active constraints via Lagrangian duality, enabling significant reductions in optimization problem size with open-source numerical solvers. We then extend this framework with certifiable safety guarantees, deriving conditions from strong convexity and Lagrangian duality that ensure all removed constraints remain satisfied and all eliminated decision variables are zero at optimality. The resulting algorithm is validated in complex urban driving scenarios using an open-source simulator based on real traffic data.The third part develops a real-time energy-aware lane planning framework for connected electric vehicles that jointly optimizes longitudinal speed and lateral lane-change decisions using vehicle-to-infrastructure communication. A hybrid approach combines short-horizon optimal control with a graph-based approximation of long-term energy costs, supported by a data-driven energy model calibrated on a physical battery electric vehicle to reason about long-horizon energy optimal decisions in real-time. Vehicle-in-the-loop experiments demonstrate over 39% reduction in motion energy consumption relative to a human driver baseline and up to 24% reduction in total energy consumption, validating both the real-time deploy ability and energy efficiency of the proposed framework.