Robot Skill Learning based on Interactively Acquired Knowledge-based Models

Industrial robots are fenced for safety reasons because of hard-programmed robot behavior, containing implicit assumptions about the environment. A knowledge base (KB), containing objects and robot skills, can be built to make assumptions explicit [1] and classic AI search-, reasonand planning-algorithms can be deployed. Advances in robotic behavior representation via Behavior Trees (BT), skills with preand post-conditions [2] and Intuitive Programming [3] result in human understandable, explainable and therefore trustworthy task executions facilitating Human-Robot Collaboration. Considering a task to navigate around an obstacle, non-expert robot programmers tend to set a way-point at the startand end-point, not considering the robot calculating a direct trajectory, resulting in a crash. This shows that despite the above-mentioned efforts, humans still imply assumptions, resulting in incomplete, imperfect and incorrect task descriptions, demanding autonomy from the robotic agent. Autonomy can be reached with a learning algorithm exploring the continuous robotic action spaces efficiently. Reinforcement Learning (RL) algorithms utilize (raw) data stored in databases to learn a behavior that solves a specific problem under uncertainties [4]. Off-policy learning can be used to learn from human demonstration examples [5]. Finding solutions in high-dimensions combined with reward-sparsity is computationally expensive. If learning is realized in an end-toend fashion, the found solutions can be non-intuitive, impairing the possibility for Human-Robot Collaboration. Related work, combining robot behavior representation through BT and learning the structure of the BT, can be found in [6]. Furthermore, the leaves of the BT can be learned, which represent, but are not limited to, robot skills [7].