Human-Centered Robotics Colloquium - Ton Juny Pina, University of Groningen

Title: Field applications in chemical sensing -- from portable safety monitors to mobile robots

Abstract:

Field applications in chemical sensing -- from portable safety monitors to mobile robots -- require models that are lightweight, fast to respond, and can be recalibrated from limited labelled data. Current odour classification pipelines assume controlled sampling cycles and offline feature extraction, and rely on models too computationally expensive for edge deployment.

With a few thousand neurons and sub-milliwatt power consumption, insects reliably classify odours in open environments. Inspired by this system, we present a neural network architecture that expands raw sensor array readings into a sparse, high-dimensional binary representation and classifies them with a single-layer Hebbian readout.

We evaluate the system on continuous gas mixtures where successive presentations overlap before sensors reach steady-state. Using exclusively binary weights and local learning rules, the architecture outperforms linear classification on both single-gas and binary-mixture datasets, achieving a linearization of the feature space comparable to non-linear kernel methods. The system reaches near-peak accuracy with fewer than 50 labelled exposures and produces reliable predictions within 5–7 sensor readings. The architecture also maintains online class-level frequency estimates, allowing the system to distinguish familiar from novel stimuli.

The resulting system is directly compatible with ultra-low-power neuromorphic hardware, offering a practical path toward autonomous, on-field chemical sensing.