HRI: Trust and Compliance

Beam Robotics

I planned, executed, and assesed different compliance behaviors with changes in the Beam robotic design—height and voice. For my researchers out there, it’s a 2 x 2 factorial experimental design! For solutions to challenges that I faced, skip to the end.

Planning

In order to disguise the true purpose of the study, to reduce unconsciously affecting the behavior of the participant,  we decided to use a sleep study as a front. The study from the viewpoint of the participant is interacting with a specialized robot that analyzes and recommends changes in sleep behavior based on the participant's disclosed sleep habits. In reality, we are changing the voice and height of the robot to see if the interaction/perception changes.

Method

We conducted a between groups 2 x 2 factorial experimental design study, where we wanted to measure how people disclose and behave around a certain robotics design.

The research method that we used is called the Wizard of Oz (WOz) technique. The WOz method leads the participants to believe that they are interacting with a working prototype of the interface, in this case a sleep diagnostic robot,  when in reality a researcher is acting as a proxy for the interface in another room/location. This enables the researchers to observe the openness and willingness of the participants to a new system. We wanted to see the behavior of the participants when interacting with a robot that gave them recommendations on their sleep. What they didn't know is that the independent variables that we changed were height and voice.

Needed to move forward:

  • Scales for disclosure- What constitutes as disclosing?

  • Success metrics for speech, gesture, posture, body language, and eye gaze

  • How do we change the height of the robot w/o manipulating its hardware?

  • Research Assistant script

  • TTS/human voice script and recordings for robot

Challenges and Solutions

Challenge: Setting up the built-in camera on the robot to only view one way was a challenge.

One of the main concerns with doing a WOz is that the researcher needs to make sure that the interface is as believably programmed/autonomous as possible. If we set up the lab and the study in a way that the participant notices early that a researcher is serving as a proxy, the behavior that we want to measure may be confounded. Because we wanted the robot to move and greet the participant, if the camera showed the researcher's face, then the experiment would not work.

Solution:

To set up a dual screen on a separate monitor so that we could see and hear the participant, while also moving and responding seamlessly on a separate laptop.

We needed 3 screens and two laptops. One laptop to maneuver the robot, view the participant, and pick up the selected responses. Another to select the robot's reaction/question. The extra monitor was for the dual screen to project the robot face and listen to the participant response.

Challenge: They could hear an echo of their voice

When the participant responded because laptop speaker # 1 was good enough to pick up the voice from the monitor.

Solution:

We tested different volumes from the monitor speaker, laptop #1 speaker, and the built-in speaker settings of the robot until the participant side could not hear the echo.

Challenge: Consistency between the text to speech and the human voice

Because we went through three to four iterations on the script, scheduling the voice actor for the human voice was difficult. We could not start the study without the voices internally consistent.

Solution:

Changing the script in the TTS program was just a matter of re-typing the response and mapping it to the right button. For the human voice, we had to contact the voice actor again to see if they could send us a recording. Because that person was unable to send us a recording, a fellow researcher recorded the whole script.

Challenge: Participant data had to throw out because they accidentally heard both the TTS voice and the human voice

Solution:

For the participant data that was thrown out, we had to recruit participants for that variable bracket and run the study again without the mistakes.

Previous
Previous

App Design: Collective Citadel

Next
Next

Sphero