CS Events Monthly View

Abstract:

Visualizing the diverse motion skills of biological agents and replicating them with robots has been a long-standing goal for artificial agents. However, even when they can match the motion performance, they require extensive training, which is computationally demanding. This is in stark contrast to the effortless emergence of motion in biological agents. In this thesis, we will present our efforts to mitigate the required computational burden by translating the architectures of evolutionarily-trained motion control networks into algorithms for realistic character animation and effective robot control. In the first part of this thesis, we will present a set of biologically plausible controllers that draw inspiration from networks of spiking neurons (SNN) found in insects and fishes. The SNN drives the locomotion of a hexapod and the crawling and swimming of an amphibious salamander while incorporating speed adaptation and dynamic obstacle avoidance. Next, we present an SNN that mimics the human oculomotor system to drive the tracking of a visual target with a robotic head comprising of two eyes and a neck. In the second part of this thesis, we will present a framework that alleviates the need for expert neuroscientific knowledge when designing SNN controllers that drive motions with desired dynamics. The framework is based on inverse generative modeling and uses an abstract genetic code to generate the SNN architectures and an evolutionary algorithm to automatically capture the joint dynamics from motion sequence samples. We also demonstrate the deployment of our biologically plausible methods on neuromorphic hardware to maximize resource efficiency. Our work highlights the effectiveness and efficiency of bioinspired methods in driving motion behaviors for animated characters and robots.