Training a robot with limited computing resources to crawl using reinforcement learning

In recent years, new successes in artificial intelligence and machine learning have been continuously achieved. However, this progress is largely based on the use of simulations as well as numerous powerful computers. Due to the volume taken up and the necessary power to run these components, this is not feasible for mobile robotics. Nevertheless, the use of machine learning in mobile robots is desirable in order to adapt to unknown or changing environmental conditions. This paper evaluates the performance of different reinforcement learning methods on a physical robot platform. This robot has an arm with two degrees of freedom that can be used to move across a surface. The goal is to learn the correct motion sequence of the arm to move the robot. The focus here is exclusively on using the robot's onboard computer, a Raspberry Pi 4 Model B. To learn forward motion, Value Iteration and variants of Q-learning from the field of reinforcement learning are used. It is shown that since the structure of some problems can be described by a very limited problem space, even when using a physical robot relatively simple algorithms can yield sufficient learning results. Furthermore, hardware limitations may prevent using more complex algorithms.