Search

In this paper, we investigate reinforcement learning model explainability through a pick and place task. Two robots with three degrees of freedom learned to solve the pick and place task in simulation as well as reality. To investigate the explanatory factors implicitly learned by the models, we derive robot parameters, i.e., the length of the robot segments. To overcome the black box nature of reinforcement learning models and provide a physical explanation of the results, the robot dimensions are derived from the learned reinforcement learning model and compared to the real dimensions. The hypothesis in the presented work is that converged reinforcement learning models must learn the robot parameters implicitly in order to learn a task. This transforms black box models into white box models, where each model’s decisions can be interpreted. Our experiments show that robot parameters can be derived from learned models and that the chosen reinforcement learning model implicitly learns physical context. In order to create robust and trustworthy AI systems for intelligent factories, we suggest that a physical interpretation of all black box models must be done. [ABSTRACT FROM AUTHOR]