A Neuromorphic SLAM Accelerator Supporting Multi-Agent Error Correction in Swarm Robotics

Abstract

Ultra-low-power SLAM has been of importance for edge devices to achieve extensive exploration under a GPS-restricted environment. Visual SLAM on edge devices suffers from accumulated odometry errors until re-localization occurs. This paper presents a neuromorphic SLAM accelerator supporting multi-agent error correction for applications in swarm robotics. The proposed multi-agent neuromorphic SLAM (MAN-SLAM) accelerator suppresses odometry errors by multi-agent map optimization. The MAN-SLAM accelerator employs time-domain spiking neural networks and emulates continuous attractor networks. The proposed MAN-SLAM demonstrates robust SLAM performance under the outdoor exploration of real environments.