SpiNNaker 2 Chip Runs Dynamic Neural Manifolds for Real-Time Robot Control

A research team led by Oskar von Seeler, Christian Tetzlaff, and Andrew Lehr has successfully implemented dynamic neural manifolds—low-dimensional neural trajectories underlying flexible behavior—on the SpiNNaker 2 neuromorphic chip. In biological circuits, sequential neural activity evolves along these manifolds, enabling animals to switch between behaviors fluidly. The team made this principle parameterizable by linking manifold geometry to specific circuit mechanisms (e.g., heterogeneous inhibition, gain modulation, transient currents). Their implementation runs in real-time, closed-loop on SpiNNaker 2, a digital neuromorphic platform designed for large-scale spiking neural networks.

To validate the architecture, the researchers deployed a robotic agent in a maze environment. Sensory feedback from the robot’s sensors dynamically modulated the manifold geometry—rotating subspaces to switch between behaviors (e.g., turning vs. moving forward) and fine-tuning trajectories within each behavior. The robot successfully navigated the maze without pre-programmed rules, relying entirely on the reconfigurable manifold dynamics. This work establishes dynamic manifolds as a feasible substrate for explainable neuromorphic computing, bridging biological neural dynamics with engineering applications. It opens pathways for adaptive, low-power controllers that mimic the brain’s flexibility.