Learning from demonstration and adaptive shared control for effective human-robot collaboration

This thesis explores human-robot shared control systems and the use of learning from demonstration methods to transfer skills from humans to robots. The initial investigation focuses on human adaptation to these systems, revealing that shared control helps humans learn tasks faster and adapt more effectively to robot behavior. Building on these findings, an adaptive blended human-robot shared control frame- work is proposed and assessed, emphasizing the importance of defining the scope of knowledge for the robot to accurately estimate human intention. This framework dy- namically adjusts control weights based on the confidence of this estimate, optimizing the interaction between human and robot. Real-time estimation of human intentions allows the robot to adjust its role, whether passive, leading, or collaborative. The framework is applied to a surgical suturing task, demonstrating its effectiveness in improving performance. By combining human precision with robotic support, the sys- tem enhances both efficiency and accuracy, underscoring the effectiveness of adaptive control systems that integrate human expertise with autonomous capabilities. Additionally, Context-based Echo State Networks (CESNs) are introduced as a lightweight solution for generating movement primitives in robotics. CESNs leverage context vectors to modulate reservoir dynamics, producing varied movement patterns with a single linear read-out weight vector. CESNs' computational efficiency, minimal training requirements, and resilience to noise and delays make them a promising learning from demonstration method for robotic applications.