NeRF-Nav: Hierarchical Neural Radiance Fields for Real-Time Robot Navigation and Obstacle Avoidance

Abstract Neural radiance fields (NeRFs) offer photorealistic scene representations but their monolithic structure and slow rendering hinder deployment for real-time robot navigation. We present NeRF-Nav, a hierarchical NeRF framework that enables real-time obstacle avoidance and path planning by decomposing large environments into a tree of local radiance fields with varying levels of detail. Our system introduces: (i) an occupancy-aware NeRF variant that jointly learns density and a binary occupancy grid for collision checking in constant time, (ii) a hierarchical allocation strategy that spawns and prunes local NeRF nodes based on the robot's exploration frontier, and (iii) a neural potential field planner that extracts repulsive gradients directly from the radiance field density without explicit mesh extraction. Evaluated on the Gibson, Matterport3D, and a custom warehouse dataset, NeRF-Nav achieves 94.6% collision-free navigation success at 18 Hz planning rate, outperforming both voxel-grid and TSDF baselines by 8–15% in cluttered environments. Our approach reduces memory usage by 3.8x compared to a single global NeRF while maintaining rendering quality (PSNR within 0.3 dB).