A Learning Strategy for Source Tracking in Unstructured Environments
Titus Appel, Rafael Fierro, Brandon Rohrer, Ron Lumia, John Wood · 2012
This chapter provides a brief overview of Q-Learning and its applications to robotics. Then, it formulates a source tracking problem where a robot is supposed to learn to navigate toward a light source. Next, the light-following problem is implemented in simulation and hardware using ROS, the Robot Operating System (ROS). In the simulation and hardware experiments, control delays and noise were encountered which affected the agent’s learning process. To overcome the effect of the control delays, the chapter presents an approach that modifies the Q-Learning algorithm by using the sensed actions to update the learning rule. This modification improves both the learning and the resulting paths of the light-following robot greatly. Controlled Vocabulary Terms learning (artificial intelligence); robot sensing systems