RoboBoom

Henny Admoni - Understanding People for Better Human-Robot Interaction

Henny Admoni, a Carnegie Mellon faculty member who leads the Human and Robot Partners Lab, discusses designing better human-robot interaction at KTH during her sabbatical with host Yolanda.

In this talk

  1. 4:42

    Assistive arms need better control

    Admoni describes wheelchair-mounted Kinova arms that help users with tasks such as getting a drink, but explains that many users have only one or two control dimensions for a six-dimensional arm. Switching between control modes can be slow: in one study, users spent about 15% of their time changing modes. Shared autonomy can combine a person’s joystick input with the robot’s estimate of their intended target.

  2. 10:40

    Gaze can help predict a user’s goal

    Admoni’s team added natural eye gaze to joystick signals, drawing on research showing that people often look ahead of their hand movements. In robot-arm food-spearing trials, users often watched the robot hand or a problem joint instead of fixating on their target; only 16% of counted fixations landed on the eventual goal. A sequential gaze model predicted goals earlier than joystick-only and simple gaze-counting methods, though noisy gaze sometimes led the robot to confidently assist toward the wrong object.

  3. 21:04

    Assistance can reduce perceived control

    A shared-autonomy setting called alpha determines how much the robot’s action influences movement, from user control at zero to robot control at one. In a study with people picking up a golf ball, greater assistance was linked to lower reported autonomy. Users could feel more in control even when their movements were less efficient, so Admoni argues that systems should account for user experience alongside task performance.

  4. 26:17

    Let robots ask for useful feedback

    Admoni also presents Inquire, a robot-learning system that chooses among four kinds of human feedback: demonstrations, labels, preferences between examples, and critiques of a performance. Using active learning, it selects the feedback type expected to provide the most information about the task. In Lunar Lander and Pizza Arrangement, Inquire learned more effectively and achieved better task performance than the compared approaches; accounting for the effort required to give feedback also changed which kind it asked for.

  5. 40:00

    Combine people, robots, and interaction

    Admoni’s examples draw on human sciences, robotics, and interaction research to shape both robot algorithms and how people experience them. She argues that considering people and society alongside the technology can produce more effective systems for everyday assistance.

Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.

More talks