BEE-SLAM: A 65-nm 17.96-TOPS/W Location-Sharing-Based Multi-Agent Neuromorphic SLAM Accelerator for Swarm Robotics

Abstract

Multi-agent (MA) simultaneous localization and mapping (SLAM) has been rigorously explored to enhance map accuracy in swarm robotics. Although centralized MA SLAM systems, which depend on a server for complex computations in map optimization, have been extensively studied, the circuit-domain approaches to decentralized MA SLAM systems are still limited due to challenges such as limited memory capacity and security vulnerabilities in wireless inter-agent data transmission. Thus, we propose a BEE-SLAM accelerator, a location-sharing MA neuromorphic SLAM accelerator inspired by bee communication for decentralized MA SLAM systems. The location-sharing-based MA error correction (MAEC) is employed to attain accurate map results without loop closure with a 94.81% reduced number of operations compared to the global map-based MA SLAM. In addition, a $7 {\times } 7$ pulsewidth modulation (PWM)-based hybrid mixed-signal/digital pose-cell (HY-PC) array with pseudo pose cells (PPCs) achieves $2.04{\times }$ energy efficiency compared to the oscillatory pose-cell array. The test chip fabricated in a 65-nm CMOS technology achieves a peak energy efficiency of 17.96 TOPS/W under $350 {\times } 450$ m outdoor exploration.