Assisting Humans in Human-Robot Collaborative Assembly Contexts through Deep Q-Learning

Collaborative robots, affectionately referred to as 'cobots,' serve to function alongside their human counterparts to help them complete a specific task. This differs from traditional systems in which the machines are set about their own jobs and are often locked behind cages so as to prevent human access in favor of safety. By removing these walls and introducing collaborative systems, a new level of versatility and productivity is opened within the contexts that they are often employed. This focus of human-robot interaction has grown in recent years, and alongside it the topic of teaching and learning from demonstration has been investigated. Several methods of implementation for this topic have been developed, and while they are potentially effective, they still have gaps in versatility. Thus, we propose a different method of robot learning from demonstrations through the employment of deep Q-networks. These networks permit effective learning not only with human demonstration data, but also with direct feedback from the collaborating user. The proposed solution is experimentally implemented in real-world human-robot collaborative tasks. Preliminary results and analysis suggest the competitive performance of the proposed approach. Future work of this study is also discussed.