Reinforcement Learning for Motion Control Using Unity and ML Agents

Resumen

The growing adoption of reinforcement learning in robotic motion control creates a need for trained engineers working at the intersection of robotics and machine learning. Even though this area is one that requires knowledge of advanced mathematical concepts such as policy optimization and deep neural network algorithms, its effective implementation depends on a system-level framework. Undergraduate students can study this abstraction inside a simulation and implement it using predefined coding libraries and tools. Two case studies are presented: a model of a racing car that learns to navigate three different racetracks and another model of a quadruped robot that learns to walk towards a target. The students develop the framework using a combination of preexisting elements such as physical environments and policy optimizers and encoding abstract concepts such as actions, observables, rewards, and training episode structures. Results show that undergraduate students can create models that exhibit emergent learning behavior by training policies.