NavHD: Low-Power Learning for Micro-Robotic Controls in the Wild

Abstract

Micro-robots are emerging as powerful tools for search-and-rescue, precision agriculture, and cooperative manipulation, where their small size and low cost offer advantages over larger robots. However, enabling autonomous navigation on these robots remains challenging due to severe hardware constraints, such as limited memory, energy, and computational power. We explore a brain-inspired learning paradigm called Hyperdimensional Computing (HDC) to equip a cheap, lightweight navigation model that runs onboard micro-robots. We present NavHD, which features an adaptive HD encoder that learns spatial representations and incorporates loss-based training for both imitation learning and off-policy reinforcement learning. Our hardware implementation of NavHD uses eight ultrasound sensors and is optimized to run on an ARM Cortex-M4 core, using only 10.2 kB of memory, 900 clock cycles and 1.1 mJ of energy per inference. Through experiments in both simulation and the real world, we demonstrate that NavHD outperforms DNN-based and prior HDC-based RL methods in obstacle avoidance by more than 2x the performance, while achieving 2-26x more superior resource efficiency.