Semantic Behaviour Tree Learning through Kinesthetic Demonstrations for Position-Force Controlled Robotic Applications

Robots' semantic understanding of their surroundings is critical to make them flexible and autonomous in decision-making. However, this semantic knowledge must be transferable to the robot, executable with commands that the robotic agent can understand, and adaptable in case of unexpected events. In this work, we propose a method that exploits Programming by Demonstration (PbD) to provide the robot with semantic Behaviour Trees (BTs). These encapsulate the Skill's semantic knowledge through Predicates and its execution through position and force primitives. Each tree can be enhanced with additional demonstrations, creating new Modes to achieve the same action. Together with a knowledge base and a PDDL planner, BTs provide the robot autonomy, flexibility and reactiveness.