Reachability Graph Learning with Reinforcement Learning Control for Robotic Tasks

Deep reinforcement learning (RL) has proven to be an effective method for learning robot control in simulation, but it suffers from sample inefficiency and lacks generalization capabilities in real-world applications. These weaknesses are often due to unrealistic assumptions, including the availability of a perfectly accurate simulator, the ability to collect vast amounts of training data, and the existence of a well-defined, single reward function. Coupling learned control policies with global automated planning for skill composition has improved performance in both simulated and real-world robotics settings. In this paper, we propose an adaptation of the Reachability Graph Learning algorithm for robotic applications. Our algorithm is able to explore and build a reachability graph directly by learning in the real world, without resetting the agent state. By discarding unnecessary information, especially when dealing with large observation spaces, we demonstrate the ability to construct a low-cost control policy that is highly efficient for real-world robotic goal-reaching tasks.