MIT Department of Electrical Engineering & Computer Science

DATE: THURSDAY, MARCH 23, 1995

TIME: Refreshments at 4:00
TALK AT 4:15

PLACE: NE43-518

Teleassistance: Using Gestures to Direct Robots

Polly Pook
Computer Science Department
University of Rochester

Abstract:

A next step in creating intelligent robots is to construct complex behaviors from the coordination of perception, action, and low-level cognition. We have developed teleassistance to investigate this integration, coordinating the perceptual and motor skills of a reactive robot with human cognition. The technique has immediate practical application to the field of human/robot interaction.

Teleassistance is a dual control strategy that combines the advantages of teleoperation and autonomous control. At the top control level, a human operator uses a gestural language to guide the robot as it performs a known task. Each gesture indicates the next action to perform and a hand-centered coordinate frame in which to perform it. For example, the operator may point to an object for reaching or preshape the hand for grasping. At the bottom level, a servo-controlled robot runs autonomously within the context indicated by the gesture.

Teleassistance offers two benefits. First, each gesture constrains the robot to a specific spatial and temporal context within the overall task. The servo routines can thereby position the robot in relative coordinates and limit the possible interpretations of feedback. This significantly simplifies computation and requires only a sparse model of the task. Second, the gestures are symbolic, conveying intent without requiring the person to literally control the robot. This alleviates many problems inherent to teleoperation, including poor mappings between operator and robot physiology, reliance on a broad communication bandwidth, intense concentration demands on the operator, and risk to the remote robot. We demonstrate these properties by showing experimental results contrasting teleassistance to teleoperation and autonomous control.

In teleassistance, gestures segment a task into a series of local contexts that a reactive robot can exploit. We show briefly how this mirrors the observed sequentiality and computational parsimony of human decision systems, to motivate its promise as a model for intelligent robots.

HOST: Prof. R. Brooks