End-to-end action model learning from demonstration in collaborative robotics

Access to advanced technology is crucial across all engineering disciplines. In the realm of industrial automation, collaborative robotics serves as a key solution, particularly for small or medium-sized enterprises facing frequent shifts in production demands. This paper introduces a Symbolic Programming by Demonstration approach to efficiently configure and operate a collaborative robotics workstation. While motion profiles (i.e., the how) are taught through the commonly used lead-through programming method, the conditions to check before the execution of a motion and its impact on the environment (the when and what, respectively) are automatically derived using visual feedback. Differently from related works, the present methodology does not require a pre-compiled domain knowledge to encode the semantic characterisation of a demonstrated action (i.e., preconditions and effects). An industrially-relevant use-case, consisting in a collaborative robotics assembly application, is introduced to validate the approach. Results show high success rates in interpreting and solving user-defined tasks (i.e., goals) as well as the capability of the method to generalise well in situations never seen during the acquired demonstrations.