University of Sussex,

Summary. One of the most popular approaches to developing bipedal walking machines has been to record the human gait and use it as a template for a walking algorithm. In this paper we demonstrate a different approach based on passive dynamics, neural networks, and genetic algorithms. A bipedal machine is evolved in simulation that when pushed walks either forward or backwards just enough to release the pressure placed on it. Just as a tango dancer uses a dance frame to control the movements of their follower, external forces are a subtle way to control the machines speed. When the machine is subjected to noise in its body’s size, weight, or actuators as well as external forces it demonstrates the ability to dynamically adapt its gait through feedback loops between its actuators and sensors.