Spiking Neural Networks for Noise-Robust Walk-in-Place Navigation in VR: An Image-Based Encoding of Positional Data

Classifying walking-in-place (WIP) states in virtual reality (VR) as a method for locomotion through extended virtual environments remains a critical challenge for enhancing immersion and interaction. This paper addresses WIP navigation by converting motion sensor data—captured from VR headsets and controllers, including head and hand movements—into 2D images for neural network-based classification. We evaluate two approaches: a traditional artificial neural network based on the multilayer perceptron (MLP) and a neuromorphic model inspired by biological principles, specifically the resonate-and-fire (RAF) neuron. Our findings demonstrate three key contributions. First, encoding sensor data into images effectively preserves movement history while reducing the complexity associated with raw coordinate processing. Second, we show that simple neural models can accurately classify WIP states in real-time, achieving up to 99.87% accuracy with an MLP and 98.57% with a spiking RAF network. Third, our comparative analysis reveals that the neuromorphic RAF model exhibits superior noise robustness and potential for greater energy efficiency. Furthermore, a user study involving 30 participants evaluated our neural network approach against a traditional head-analysis WIP baseline and Teleport locomotion. While Teleport was preferred for overall navigation ease, our method demonstrated significantly improved classification performance and was clearly preferred by users over the baseline head-analysis WIP method validating its effectiveness as an enhanced WIP solution. We conclude that combining imagebased sensor encoding with efficient neural classifiers offers a robust solution for WIP state discrimination in VR, with potential extensions to other motion-based interaction tasks.