Real-Time Motion Planning: From Optimal Primitives to Cooperative Negotiation and Safety Validation

Autonomous vehicles represent a fundamental shift in transportation technology. They promise to improve road safety, optimize traffic flow, and increase energy efficiency. A main motivation for this change is the goal of reducing road accidents caused by human error. Meeting this goal requires vehicles that can make safe and efficient decisions in real time across the full range of conditions they encounter on the road. This ability rests on the motion planning system. A motion planner must generate dynamically feasible trajectories in complex, rapidly changing, and non-convex environments, within strict computational limits. This requirement exposes a fundamental trade-off between solution quality and computational speed. Traditional optimal control methods produce high-quality trajectories but are too slow for online use. Sampling-based methods operate faster but often yield suboptimal or infeasible maneuvers. This thesis addresses this trade-off through the concept of motion primitives, trajectory segments derived from vehicle-specific dynamic models. Motion primitives allow a planner to rapidly construct complex, dynamically feasible maneuvers, either by drawing on precomputed offline libraries or by exploiting analytical shortcuts that make their online computation fast. These primitives are integrated into the \MPTree{} framework, a modular planning architecture that combines a sampling-based search with parameterizable maneuvers. Decoupling the search strategy from maneuver generation allows the architecture to generalize across diverse driving contexts, from comfort-oriented urban navigation to time-optimal racing. The framework is then extended to multi-agent scenarios through vehicle-to-everything communication. In this setting, the planner uses communicated desired trajectories to enable decentralized, emergent negotiation among vehicles without explicit protocols, a strategy validated in simulation and on a connected vehicle platform. A high-performance planning system introduces a corresponding challenge: ensuring its safety. Traditional validation methods, such as real-world road testing, are insufficient for capturing the rare and critical events required to validate complex, optimization-based planners. This thesis addresses this challenge with a data-driven framework for safety assessment. The methodology creates a surrogate model of the autonomous system's behavior using an ensemble of bootstrapped neural networks. This model predicts safety outcomes and associated uncertainties for millions of scenarios in seconds. It enables both large-scale offline risk assessment and online risk monitoring during vehicle operation. High-performance racing places additional demands on the underlying motion primitives, requiring velocity profiles that are both time-optimal and computed at speeds suitable for online use. The Forward-Backward with Generic Acceleration (FBGA) algorithm addresses this need. FBGA computes optimal velocity profiles along fixed paths under complex, non-convex acceleration constraints. By adaptively determining acceleration and braking switching points instead of relying on apex pre-computation, it achieves accuracy comparable to offline optimal control while being up to three orders of magnitude faster. FBGA is further deployed on a physical racing platform, both as a real-time velocity replanner and as a trajectory evaluation tool within a sampling-based planner. Together, these contributions share a common methodological foundation in motion primitives, applied across complementary problems in autonomous driving: a flexible, general-purpose trajectory planner extended to cooperative multi-agent settings, a scalable methodology to validate its safety, and a fast, accurate velocity-planning algorithm for high-performance driving. This body of work contributes a set of methods toward safer and more capable autonomous driving systems.