NEUROMORPHIC LIDAR-BASED ROBOT NAVIGATION USING SPIKING NEURAL NETWORKS AND COMPUTE-IN-MEMORY PROCESSING

Abstract

Autonomous navigation is one of the core capabilities that enables robots to navigate in real-world environments by continuously making decisions without human intervention and manual control. It requires reliable obstacle avoidance under strict latency and energy constraints. Deep learning-based methods such as deep neural networks can effectively map sensor readings to controlled outputs but deploying them directly onto embedded robotic platforms is challenging due to computing and memory overhead. This thesis investigates non-Von Neumann computing i.e., neuromorphic computing using 2D LIDAR, combining Spiking Neural Networks (SNN) for event-driven processing with memristor-based analog compute-in-memory for hardware aligned inference. The work is organized into two experimental studies. Study 1 examines how Leaky Integrate-and-Fire (LIF) spiking neurons activation model in an SNN architecture impacts SNN dynamics and navigation performance in LIDAR-based obstacle avoidance tasks. A supervised dataset has been collected using a small differential wheeled robot, consisting of 59x59 spiking frames derived from LIDAR scans and paired with robot’s internal kinematic state and motion control commands. A Convolutional Neural Network (CNN) has been used as a baseline model to compare the results predicted by an SNN model with surrogate gradient training using different values of the leakage constant α ranging from 0.1 to 0.9. The results show that with an optimal value of α, the proposed SNN can achieve the same level of performance as CNN, offering low operations counts. Study 2 demonstrates a real-time, closed-loop navigation pipeline using an 8×8 memristor crossbar integrated onto an OpenMENA board with a large Ackermann-steered robot. The crossbar performs first-stage analog inference using programmed conductance states. A lightweight Multi-layer Perceptron (MLP) then refines the noisy hardware readout to compensate for non-idealities. Robot trials validate the feasibility of memristor-in-the-loop inference for generating stable control signals in real time. Overall, this thesis shows that SNN navigation performance is highly sensitive to leakage dynamics and memristor compute-in-memory can be practically embedded in a robotic control loop coupled with hardware-aware refinement, supporting the broader viability of neuromorphic approaches for efficient LIDAR-based robot navigation.

Date of Award	2026
Original language	American English
Awarding Institution	HBKU College of Science and Engineering