Optimized Local Path Planner Implementation for GPU-Accelerated Embedded Systems

Autonomous vehicles are latency-sensitive systems. The planning phase is a critical component of such systems, during which the in-vehicle compute platform is responsible for determining the future maneuvers that the vehicle will follow. In this paper, we present a GPU-accelerated optimized implementation of the Frenet Path Planner, a widely known path planning algorithm. Unlike the current state-of-the-art, our implementation accelerates the entire algorithm, including the path generation and collision avoidance phases. We measure the execution time of our implementation and demonstrate dramatic speedups compared to the CPU baseline implementation. Additionally, we evaluate the impact of different precision types (double, float, half) on trajectory errors to investigate the tradeoff between completion latencies and computation precision.