New SNN on SpiNNaker2 achieves 10,000x energy savings for trajectory control

Neuromorphic controllers for size-, weight-, and power-constrained systems often struggle with interpretability or fail to fully exploit neuromorphic dynamics. Nourse and Quinn break this trade-off with a new spiking neural network (SNN) architecture that uses a winner-take-all (WTA) circuit augmented with accessory populations to induce controlled transitions in neural activity. By employing shunting inhibition, the network decouples the control of direction, speed, and radius in polar trajectories, making the system both interpretable at the system-dynamics level and highly efficient. The architecture is analytically tunable, offering clear design rules for engineers.

The team implemented the SNN on the SpiNNaker2 neuromorphic processor and compared its performance against conventional computing platforms. Results show a two to three orders of magnitude reduction in wall-clock step time (100–1000x faster) and a three to four orders of magnitude reduction in energy expenditure (1000–10,000x less energy). This leap in efficiency opens the door to low-latency, energy-efficient control loops for drones, rovers, and other autonomous robots operating at the edge. The work combines theoretical rigor with practical deployment, setting a new benchmark for neuromorphic control in robotics.