Elimination of Race Condition During GPU Acceleration of Probabilistic Height Map

Abstract

Walking robots or mobile robots need local map data for generating versatile motion. Especially, 2.5D probabilistic height map is being viewed as a robust method that reduce computational cost and memory usage. This paper attempted to accelerate a probabilistic height mapping algorithm using GPU instead of CPU, and propose an algorithm for solving a "race condition" problem that arises during generating height map data by the GPU thread block. The accelerated algorithm achieves that all the depth data incoming 30Hz is successfully processed into the height map data, and theoretically guarantees the minimum idle time of the thread block. Finally, its feasibility is also verified on simulator that includes a footstep generating algorithm and a NMPC (Nonlinear Model Predictive Control) walking controller.