Computer Science Seminar - Dr. Lea Steffen, University of Waterloo (Ontario)

Title: Walking, Drawing, Deciding: Coordinating a Humanoid Robot with Spiking Neural Networks

Abstract:

One of the many tasks of our brains is to coordinate whole-body movement, and they perform this function with an impressively low energy footprint. Robots that perform the same function usually rely on conventional neural networks, which are computationally expensive. Spiking neural networks, which communicate through brief pulses like real neurons do, running on neuromorphic hardware could bridge this gap. However, prior work has mostly focused on controlling either the arms or the legs in isolation and has rarely involved a full-size humanoid. In this talk, I will present a spiking architecture that coordinates both arm and leg movements. A model of the basal ganglia, the brain region responsible for action selection, determines whether the robot should walk or move its arm. A brain-inspired arm controller reaches for targets and traces handwritten digits, while a spiking version of a standard walking policy handles locomotion. In simulations, a full-size Unitree H1 humanoid robot follows a walking path, draws digits and switches between the two tasks, with spiking neurons controlling the arm, the walking, and the choice between them.