REN: A Reconfigurable End-To-end NeuroSLAM Hardware Accelerator for Micro Mobile Robots

Visual Simultaneous Localization and Mapping (VSLAM) is essential for enabling autonomous systems to navigate in unknown environments. Brain-Inspired VSLAM algorithms are valued for their energy efficiency and robustness. Inspired by the hippocampal mechanisms in rodents, this paper proposes a Reconfigurable End-To-end NeuroSLAM (REN) accelerator for micro mobile robots. The proposed end-to-end hardware architecture, with a novel Dynamic Reconfigurable Dual-Mode-PE Cluster (DR-DPC), is employed to fully accelerate the NeuroSLAM algorithm to reduce energy consumption and processing latency. The DR-DPC with 8 PE groups is proposed to accelerate the three major tasks of the NeuroSLAM, to save hardware overhead and achieve high hardware utilization. A Dynamic PE-group Allocation Scheme (DPAS) for the DR-DPC is used to efficiently allocate the PE groups to balance the latency of visual odometry and loop closure detection tasks to improve hardware utilization, reduce the process latency and improve energy efficiency. Hardware implementation results at 40 nm and 1.1 V exhibit that the proposed REN accelerator design can achieve a high energy efficiency of 1.04 TOPS/W, a PE utilization of ~85% in average, and a frame rate of 90 fps@VGA at 120MHz, among the state-of-the-art designs.