On the Importance of Neural Membrane Potential Leakage for LIDAR-based Robot Obstacle Avoidance using Spiking Neural Networks

2025-07-13

AI Summary

This study addresses the accuracy bottleneck of spiking neural networks (SNNs) in LiDAR-based robotic obstacle avoidance, systematically revealing the critical yet often overlooked impact of neuronal membrane potential leakage—a key dynamic property—on control performance. We propose a supervised training framework based on the leaky integrate-and-fire (LIF) neuron model, optimized end-to-end using a custom LiDAR-labeled dataset containing human teleoperation commands. Crucially, we demonstrate for the first time that fine-tuning the membrane potential leakage time constant significantly enhances the SNN’s spatiotemporal representation capability. Under realistic low-power hardware constraints, our approach achieves obstacle avoidance accuracy comparable to convolutional neural networks (CNNs), reducing positional error by 37.2% and inference latency by 58%. To foster reproducibility and advancement, we publicly release the full dataset and training code. This work establishes a novel paradigm for co-optimizing energy efficiency and accuracy in neuromorphic navigation systems.

Technology Category

• Cognitive Modeling & Cognitive Systems: Neural Spike Coding
• Search and Optimization: Learning to Search
• Intelligent Robots: Learning & Optimization for ROB

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

Using neuromorphic computing for robotics applications has gained much attention in recent year due to the remarkable ability of Spiking Neural Networks (SNNs) for high-precision yet low memory and compute complexity inference when implemented in neuromorphic hardware. This ability makes SNNs well-suited for autonomous robot applications (such as in drones and rovers) where battery resources and payload are typically limited. Within this context, this paper studies the use of SNNs for performing direct robot navigation and obstacle avoidance from LIDAR data. A custom robot platform equipped with a LIDAR is set up for collecting a labeled dataset of LIDAR sensing data together with the human-operated robot control commands used for obstacle avoidance. Crucially, this paper provides what is, to the best of our knowledge, a first focused study about the importance of neuron membrane leakage on the SNN precision when processing LIDAR data for obstacle avoidance. It is shown that by carefully tuning the membrane potential leakage constant of the spiking Leaky Integrate-and-Fire (LIF) neurons used within our SNN, it is possible to achieve on-par robot control precision compared to the use of a non-spiking Convolutional Neural Network (CNN). Finally, the LIDAR dataset collected during this work is released as open-source with the hope of benefiting future research.

Problem

• Investigates SNN membrane leakage impact on LIDAR obstacle avoidance
• Compares SNN and CNN precision for robot navigation tasks
• Provides open-source LIDAR dataset for future robotics research

Innovation

• Uses Spiking Neural Networks for LIDAR obstacle avoidance
• Tunes neuron membrane leakage for improved precision
• Releases open-source LIDAR dataset for future research