Real-World Projectile Catching with Reinforcement Learning: Empirical Analysis using Discretized Simulations

Robotic projectile catching has previously been done with path planning, in which the trajectory is predicted using basic kinematic equations. In addition, reinforcement learning (RL) has proven successful in mastering video games, such as 8bit Atari 2600 games like PacMan or Breakout. We evaluate the performance of two RL algorithms, Q-Learning and a Deep Q-Network (DQN), applied to a simulation of a novel application of catching a projectile. A continuous physical environment is translated into a discretized, limited state and action space, and then solved using RL.