Reinforcement Learning for Fault Tolerant Robots by Learning State Transition Differences

Abstract

Robots have the possibility of breaking down, and when in an environment where access is limited, they still need to accomplish tasks they are required to do, even when reparation is not a possibility. The purpose of this research is to derive a policy using reinforcement learning that produces a high-performance robot, even in the case of failure. The proposed method learns the normal transition function and adds the difference between the predicted state transition and the actual state transition to the input of the policy network. The results of the experiment show that our method outperformed the baseline method that uses no state transition differen.