Controller design for HRI

From RSN
Jump to: navigation, search

For robots to become useful in tasks that involve, interacting and sharing physical space with humans, they will need to learn to operate in such a way that they do not disturb or harm people. We study the problem of developing robots that consider the affective state of humans during interaction. There are several questions involved with this problem. The robot has two possibly conflicting objectives: first, complete a physical task (such as clean the floor, drill a hole, etc.), and second, do so in a way that does not disturb the human participant. We study this problem in the context of a simple tracking task. The physical task of the robot is to maintian a specified distance from the human.


We evaluate the "human-friendliness" of a particular controller by measuring the human's biometric data. Instead of utilizing existing controllers designed without regard for the human affective state, we utilize biofeedback as a means of evaluating the quality of controllers during the design process. If we can mathematically relate the comfort level of the human to robot actions, then we can form an optimal control problem. Reinforcement learning or the Calculus of variations can be used to solve such problems. However, we do not know the relationship between stress response and system state, and learning the relationship through repeated trials may be to costly.


Due to the difficulty in modeling the human stress processes, we simplify the problem as follows: for a fixed human trajectory and an initial position of the robot, find a robot trajectory, (rather than a policy over the entire state space) which completes the geometric task while keeping the stress of the human below some maximum value. Our approach to the problem can be described as follows: First compute a set of controllers, each of which is optimal for some objective. The crucial point here is that these objectives do not involve human emotions and therefore corresponding policies can be computed off-line. Our technique involves stitching together portions of the geometric controllers to obtain a human-friendly controller. We repeat the process until a set of admissible trajectories is obtained.

In the tracking task, the robot is charged with maintaining a specified minimum and maximum distance from the human. The goal of the robot is to move so that it maintains the appropriate distance while simultaneously minimizing the measured stress. This task is motivated by scenarios in which the robot may be required to stay close to the human. For example this may be done in order to track the human's gesture. A galvanic skin response (GSR) sensor is used to estimate the stress of the human participant.

Paper In Journal of Autonomous Robots : Special issue on Socially Assistive Robotics