Robotic systems hold tremendous potential to assist humans not only in industrial environments but also in everyday tasks. However, effectively deploying these systems typically requires programming expertise, creating a significant barrier for novice users. Learning from Demonstration (LfD) offers a promising alternative by allowing users to teach robots through demonstrations rather than writing sophisticated code, thereby promoting the democratization of robotics. To enhance the efficacy of LfD, it is essential to understand not only the underlying learning algorithms but also the quality and information content embedded in human-provided demonstrations.
First, we quantified the quality of human demonstrations using two motion-related features: manipulability and joint-space jerk. We then investigated how these quality measures influence the performance of LfD models. Our results show that demonstration quality directly transfers to the learned models, highlighting the critical role of human-provided data in shaping robot performance. While most existing LfD approaches focus primarily on successful demonstrations, emphasizing what to imitate, failed and suboptimal demonstrations can provide valuable information about what not to imitate. To address this gap, we proposed a novel learning algorithm that not only learns from successful demonstrations but also explicitly repels the model from failed ones. This approach enables more robust and efficient learning even when demonstrations are imperfect, as commonly occurs with novice users.
Next, we examined the role of kinematic redundancy in LfD through a user study involving 24 novice participants. Specifically, we analyzed how the robot’s degree of redundancy, reflected in the number of available joint-space solutions, affects both human teaching behavior and robot performance. Our findings indicate that reduced redundancy makes it more difficult for novice users to identify desirable joint configurations, negatively impacting both the quality of demonstrations and the resulting learned models. Finally, we developed an augmented reality (AR)-based real-time feedback system to assist users in providing higher-quality demonstrations. By offering intuitive visual guidance during teaching, the system supports more effective human-robot interaction and improves overall learning outcomes.
Overall, this research advances LfD from a purely algorithmic perspective toward a more human-centered framework. By quantifying demonstration quality, leveraging failed attempts, analyzing the impact of redundancy, and introducing interactive feedback mechanisms, this work contributes to the development of intuitive and accessible human-robot interaction systems in which everyday users can teach robots naturally and effectively, expanding the possibilities for real-world, user-centered robotic applications.