Neural Network Driven Locomotion of General Bodies
Examensarbete för masterexamen
Complex adaptive systems (MPCAS), MSc
Producing locomotion controllers for general bodies has multiple applications, foremost in robotics. Most robot controllers are today developed through the use of autonomous control, trying to deviated as little as possible from some target gait. This target is usually supplied in the form of motion capture data or determined by a human. This work attempts to instead use machine learning methods to allow simulated agents to learn locomotive behaviour. The environment is void of any preconceived notions regarding what the resulting gait should look like, and simply encourages the agents to maintain forward velocity. The controllers are represented in the form of neural networks mapping sensory inputs to actuator outputs. Three machine learning methods are explored for training these networks; evolutionary algorithms with niches, the reinforcement learning algorithm Proximal Policy Optimization, and a novel algorithm combining the two previous methods. The hybrid algorithm out-performs the rest. It works well on simpler problems such as a 2-dimensional biped, but struggles on the difficult unstable 3-dimensional problem of a full humanoid robot. The results are not competitive with traditional autonomous control results when controlling well understood bodies like humanoids. However the algorithm does succeed in finding creative gaits for less structured problems and odd bodies, something that a human might find difficult.
Fysik , Physical Sciences