Behavioral Entropy in Human-Robot Interaction

摘要

The ability to quickly and accurately measure how various design decisions affect human workload is an important need in human-robot interaction (HRI) and other HMI domains. Although various techniques allow workload to be estimated, it is important to develop methods for obtaining workload estimates objectively and in real-time without interfering with the normal operation of human-robot interaction. In this paper, the authors develop behavioral entropy as a technique for estimating human workload in HRI domains. They develop their theory and then present three case studies that help validate the power of behavioral entropy. The first two studies use average behavioral entropy and lend support to the thesis that behavioral entropy discriminates between good and bad operating conditions. The third case study uses a real-time version of behavioral entropy to learn proper force feedback. This case study uses a reinforcement learning technique to show that real-time estimates of behavioral entropy are informative. Case study 1 compared the usability of two teleoperation schemes, manual teleoperation and shared-control teleoperation. Users were asked to drive a robot with their right hand using a joystick and to answer multiple-choice addition and subtraction problems with their left hand. Case study 2 compared the usability of two interfaces in robot teleoperation. The first interface displayed sensor readings in a side-by-side format. The second interface integrated these sensor readings in a pseudo-perspective view, with a representation of the robot displayed in the view. Subjects were asked to teleoperate a simulated robot through three mazes while performing a memory task. Case study 3 used a real-time estimate of behavioral entropy as a major factor in constructing an estimate of driver workload.