Robotic Learning from Successful and Failed Demonstration

Abstract

In remote applications such as space exploration, intelligent assistance robots are supposed to work autonomously without human intervention. It is essential that such assistance robots provide a large skill catalog. However, programming new actions manually is cumbersome. Instead, it would be much more efficient to acquire new skills through learning from demonstration. When training new skills, errors may occur, as some tasks are challenging to execute - even for humans. From this, a fundamental question arises: Is it possible to leverage failed demonstrations for learning instead of discarding them? To answer this question, this thesis evaluates three algorithms toward this possibility. First Gaussian Mixture Regression (GMR), second Hidden Markov Model (HMM), and third Dynamic Movement Primitive (DMP). These algorithms are designed to generate the trajectory of a robot’s end-effector depending on the pose of a target object. Using this approach, successful and failed demonstrations are applicable to be used to learn a versatile policy. The methods are compared and used as a geometrical description of a task, where the target pose of the end-effector is optimized with respect to the best possible reachability. The on-line motion generation is then integrated into the cognitive reasoning framework on the humanoid robot Rollin’ Justin. The different motion generation methods are evaluated and experiments on the humanoid robot prove the successful application of the proposed method. In conclusion, this thesis shows that it is possible to learn explicitly from failed demonstrations if the symbolic information is known.