Robust Segmentation and Classification of Robot Skills from Demonstrations with Detection of Unknown Skills

The shift towards more customizable products drives small and medium-sized enterprises to adopt flexible production systems. The ability to quickly reconfigure robots for new tasks is crucial, highlighting the need for easy programming tools due to the scarcity of robotics experts. This paper presents a segmentation and classification method to autonomously generate from a kinesthetic demonstration the sequence of robot skills required to deploy the robot for a task. The approach begins by describing the environment using predicates and defining robot skills by their preconditions and effects. Demonstration segmentation is achieved by detecting transitions in the world state, and segments are classified using a cyclic skill-guessing algorithm. The method is validated with a skill library that includes manipulation and tool skills such as welding and polishing. Results show performance comparable to the baseline with improved robustness to missed predicate activations and the ability to identify missing skills, prompting operator intervention.