Dongheui Lee - Human Centric Approaches for Robot Learning and Interaction with Human

In this talk

1. 0:46

   Learning from demonstrations

   Lee describes teaching robots through imitation, physical guidance, and teleoperation, including with haptic devices and VR. Examples range from sweeping and making drinks to sorting objects by properties such as size or weight.

2. 2:20

   Learning physical collaboration

   Robots can learn contact-based interactions such as responding to a person’s gesture with the appropriate high five. In shared transport, a robot can progress from following passively to anticipating a human’s intended route; in polishing, it can adjust how control is shared when unexpected forces occur.

3. 5:38

   Using context to anticipate intent

   Lee says real-world deployment remains difficult because settings are dynamic, safety matters, and robot data is limited. Her group estimates intent by considering human-object interactions and combining language, visual, and spatial cues, rather than relying only on a person’s movement or destination.

4. 9:13

   Assist before the person asks

   In a kitchen example, the robot anticipates that someone near a laptop may next reach for a cup, so it prepares to pour. If the person switches to work, the robot returns to idle; it waits for the person to hold the cup before pouring. Another demonstration shows a robot handing over bottles or a book based on the person’s activity.

5. 11:52

   Infer goals, not just likely patterns

   Lee describes this research as early-stage and argues that predicting frequent patterns is not enough. Robots should infer why a person is acting in a particular way, then use that goal and the situation to anticipate possible tasks and plans. She also notes that everyday human video can help teach interactions, alongside robotic data.