GPU-Accelerated Differential Evolution for Robotic Planning and Control

We consider efficient, real-time trajectory optimization for robotic planning and control tasks with complex, non-convex dynamics and constraints. We propose a general-purpose CUDA-accelerated optimization framework that parameterizes control trajectories using Bézier curves and evolves them using SHADE, a state-of-the-art adaptive differential evolution algorithm. This enables efficient, implicit search in hybrid control spaces that are high-dimensional, non-convex, and non-differentiable. Our approach directly decodes dynamically feasible trajectories from compact Bézier parameters via analytical models, allowing cost evaluation over entire continuous rollouts. We exploit GPU parallelism to accelerate population evolution and batched trajectory simulation. We demonstrate the method on three challenging domains: underactuated cart-pole swing-up, humanoid locomotion with discrete footstep decisions, and dynamically constrained autonomous vehicle planning. Across all domains, our framework discovers feasible and high-quality trajectories significantly faster than CPU-based solvers, while maintaining task generality, model flexibility, and real-time applicability.