Neuromorphic Place Recognition

About the project

Two grants from the Intel Neuromorphic Research Community supported this work: Spike-based Visual Place Recognition using Intel’s Loihi and Active, adaptive and predictive robot localization.

Various feature descriptors have been proposed for conventional place recognition, with different advantages ranging from global techniques that are highly viewpoint invariant to local techniques that achieve superior precision. However, it was unknown whether these properties continue to exist when processed in a sparse manner suitable for spike-based processing on Loihi.

The follow-on project tackled open problems that are particularly well suited to the bottom-up promise of neuromorphic computing, namely continual, adaptive learning suitable for power-constrained closed-loop robot control. Place recognition needs both “slow” adaptation (e.g. urban vs rural, day vs night) and “fast” adaptation (e.g. sudden onset of rain, entering tunnels).

Outcomes

The work produced some of the first spiking neural networks for visual place recognition, scaled them from small regions to entire cities with ensembles of compact networks, and led to LENS (Science Robotics, 2025), a compact neuromorphic system for ultra energy-efficient, on-device robot localisation. I also contributed to NeuroBench, the community framework for benchmarking neuromorphic computing algorithms and systems.