| Summary: | Although recent advances in robotics enable the automation of manual tasks in manufacturing, integrating robots into a factory remains time and resource intensive, as it requires conventional robot programming and robot experts. In order to increase the feasibility of robot integration into industrial processes, the programming of robots must be easily accessible to domain experts with little to no experience in robotics. In this thesis, we present Active Keyframe Learning (AKL) for learning the task specification as an ordered sequence of keyframes to capture the physical interactions and geometric constraints from a single demonstration of a task given by a nonexpert. We learn the least restrictive task specification that maximizes the flexibility given to a motion planner by learning the human intent for demonstrated constrained motion online and performing interaction-based and constraint-based segmentation offline. We conduct a user study to evaluate the keyframe, pose, constraint accuracies, workload, and teaching efficiency of AKL against two state-of-the-art techniques in keyframe and constraint learning and demonstrate the significant benefits of utilizing AKL to teach tasks to robots.
|
|---|