Device Design for Natural Human-Robot Interaction in Collecting Demonstration Data for Robot Learning

Good manipulation learning depends on recording the right data signals. The human hand is guided not just by where it is in space, but by where it touches an object and how hard it is pressing, yet most demonstration systems capture only motion and camera data, leaving out the contact and force information that often matters most. Worse, the recording hardware itself frequently gets in the way, changing how a person naturally handles an object. This thesis describes the hardware and data-collection infrastructure we built toward natural, tactile-rich demonstration, as part of a broader effort on intent-aware robot learning in visually occluded settings. Three things came out of it. First, we analysed an open-source tendon-driven hand and developed a kinematic and tendon-transmission model, which led to two mechanical fixes: a chamfered winch that stopped the tendon from slipping and a thicker, double-jointed coupling link that stopped the fingers from breaking. Both worked, though extended use made clear the platform itself is too fragile to build on, and we document what a successor hand needs to avoid. Second, we built a low-cost glove with force-sensitive resistors on both sides of the hand, with a calibration model and a pipeline that merges the contact data with motion-capture data from a Rokoko glove. Third, I designed a wireless LED band that sits on the robot’s last joint and shows the operator which control mode is active,the mode sleeve,and ran a small user study comparing it against a screen-only baseline. Experienced operators preferred having the sleeve; it let them keep their eyes on the end-effector rather than glancing at the screen, though the benefit depended on having already learned the colour encoding. Together these three components form a toolchain for collecting demonstrations naturally, and for transitioning from human demonstration to autonomous execution with teleoperated fine-tuning.