BatchRRT: A Fully-Batched Reformulation of RRT With GPU-Accelerated SDF Collision Checking for Real-Time Human-Robot Collision Avoidance

Abstract

Industrial human-robot collaboration requires real-time motion planning with accurate collision avoidance in dynamic environments. While recent neural signed distance field (SDF) methods achieve sub-millisecond inference from sparse observations, traditional RRT planners fail to exploit this capability due to serial execution and linear nearest-neighbor search bottlenecks. We present BatchRRT, a fully-batched reformulation of RRT that jointly optimizes batch sampling, KD-tree acceleration, and GPU-parallel SDF collision checking. Through comprehensive experiments across 5340 planning trials in human-robot interaction scenarios spanning single and dual-person configurations, we demonstrate 4-5× speedup over baseline RRT with consistent 99-100% success rate. Our key finding reveals that with fast neural SDFs, nearest-neighbor search—not collision detection—becomes the dominant bottleneck, with KD-tree optimization providing 15.6× acceleration. Systematic ablation studies show that these three components exhibit strong interdependence: isolated batch sampling provides only modest gains (1.1×), while their combination unlocks substantial improvements. We validate our approach on a 6-DOF industrial manipulator navigating around dynamically posed human obstacles in both reaching and bending motions, including dual-person scenarios.