Evolving locomotion for virtual quadrupeds

Abstract

Determining efficient gaits and walk-cycles for arbitrary body shapes is an ongoing problem that has a wide array of applications, from robotics to video game development and computer animation. Many different methods have been used in solving this problem, each with trade-offs in run-time efficiency, generality, and ease of implementation. The technique used in this project is Proximal Policy Optimization, a form of reinforcement learning in which efficient walk cycles can be learned and improved automatically. This technique will be applied to a quadrupedal agent, which will learn to walk to a target location in a simulated environment. In addition, this project further optimizes the body of the agent over time for more efficient locomotion with genetic algorithms. In each generation 10 randomly mutated quadruped agents will be created, their performance evaluated, and the performance evaluations used to produce the next generation. In this way the agent’s body and gait will evolve together to achieve the desired results.