Hybrid Robotic Reinforcement Learning for Inspection/Correction Tasks

Abstract The ability to rapidly program robots for complex tasks is an important precursor to wider adoption of robotics in industry. Robot programming is often time consuming and brittle to unanticipated variations in processing. Automated robot task learning is a solution to this problem. Reinforcement Learning (RL) is a commonly used approach for a robot to autonomously learn simple tasks. In RL, rewards are used to guide the robot towards learning an optimal plan or control policy. RL, however, has proven to be of limited value for problems with large-state spaces and considerable environmental variability. In this paper, we investigate formulation of the RL approach for inspect/correct types of tasks, specifically a misplaced block in a simple grid-world environment (requiring searching the gird world to identify a missing block and returning the missing block back to the target). We use a hybrid method, combining the SARSA algorithm and a model of the environment. The model of the environment is used as a reference model to reduce the state space, avoiding unnecessary exploration of the environment. A main focus of this research is the impact of task variability on RL performance.