A research team from the Queensland University of Technology (QUT) has developed a novel approach to robotic navigation inspired by the neural mechanisms of insects and animals. The study, led by postdoctoral research fellow Somayeh Hussaini and supported by Intel, was published in the IEEE Transactions on Robotics journal. The team included Professor Michael Milford and Dr. Tobias Fischer from the QUT Centre for Robotics.
The research introduces a place recognition algorithm based on Spiking Neural Networks (SNNs), artificial neural networks designed to emulate the discrete signal processing of biological brains. According to Miss Hussaini, SNNs are well-suited for neuromorphic hardware, which mimics biological neural systems, offering advantages in processing speed and energy efficiency.
Modern robots often encounter challenges in navigating complex environments and rely on energy-intensive AI training. In contrast, animals demonstrate highly efficient and robust navigation capabilities in dynamic settings. Dr. Fischer noted that the research aims to create biologically inspired systems that could eventually rival or exceed conventional navigation methods.
The QUT team’s system uses modular neural networks to identify specific locations through visual input. By employing ensembles of SNNs and analyzing sequences of images instead of individual frames, the system achieved a 41% improvement in place recognition accuracy. This approach enables adaptability to changing environmental conditions, such as variations in weather and seasons.
The system was tested on a resource-constrained robot, demonstrating its practicality in scenarios where energy efficiency is paramount. Professor Milford highlighted the system’s scalability and potential real-world applications. Miss Hussaini emphasized its relevance for energy-critical domains, including space exploration and disaster recovery, where operational efficiency and rapid response are essential.
Photo: Queensland University of Technology
