An adaptable controller for autonomous robots operating in unexplored environments
GARY T. ANDERSON, Youlong Yang, Gang Cheng · 2005
Several oscillator-based robot controllers were tested for effectiveness in exploring a dynamic environment. The goal of the project was to have a team of three robots explore an arena with variable lighting and obstacle patterns and, as a team, absorb as much light as possible. Reinforcement learning was used to optimize the parameters of a coupled nonlinear oscillator controller for operation in an unknown environment. The performance of this controller was compared to those of a reactive-based controller, an oscillator controller with fixed parameters and an oscillator controller that used sensor-based parameter adjustments. Experiments indicated the reinforcement learning controller offered better performance than the other controllers. The drawback to using reinforcement learning is that a learning period is required for each milieu the robot operates in. If it is necessary to quickly explore an unknown area or a rapidly changing environment, then the controller with a sensor-based parameter adjustment has an advantage