Shiverbot - CMU Human-Robot Interaction Research Study


How long will people stay to “warm” a robot if we imbue the robot with the bodily function of shivering (versus not shivering)?

By posing the question, “how long will passersby manually turn a crank to ‘warm’ a robot if we imbue the robot with the human characteristic of shivering (versus not shivering)?”, we hoped to reach beyond the surface-level interactions taking place in order to better understand how visual confirmation of biological functions affects empathy.

Our proposed Human-Robot Interaction experiment, “Shiverbot,” explored how humans’ empathy varies in response to tangible expression of biological functions. We chained a utility cart to the railing next to where the Pausch Bridge meets the entrance of Purnell and place our custom-built, knee-height robot on that. On the cart next to the robot would be a box resembling a small generator with a hand crank attached to it. 

We were interested in finding out (1) how much time the user is willing to spend helping a robot solve a biological problem it cannot truly experience and (2) if the amount of time differs based on whether or not the user sees the robot shivering. If the amount of time users spend helping the shivering robot as opposed to the static robot differed by a statistically large enough margin and if our data sample possessed statistical validity and generalizability, we could have inferred that visual demonstration of bodily functions does impact empathy. This experiment provided greater depth in understanding how closely humans related themselves to robots and whether humans will ignore the disjunction between robots and biological functions in favor of their innate empathy.




Study Design


In the first meeting, we discussed various types of research studies involving Reliability, Empathy, and Trust in HRI. We then went on to conduct literature reviews in each of those areas. After coming together and getting feedback for each of the research ideas, we decided on making ShiverBot because embueing a robot with human bodily functions to elicit empathy has been rarely seen in the field of HRI.




Plan & Procedure

Shiverbot would actively try to initiate contact with passing individuals by saying “Sorry! Excuse me?” We wanted to avoid the situation where a group of people gathers around Shiverbot. To do so, we will avoid initiating contact when there is a large number of incoming people.

Once someone gets closer, we followed our script and provided the problem and backstory of Shiverbot. If the person attempts to leave early, we will respond with “Really?” followed by lines from contact termination. We also made a list of possible user responses and ways we would respond as Shiverbot.

For contact termination, we decided to ask for the name of the participant in order to make them feel as though they made a difference in the robot’s life.




Measurements & Observations

Started from the point of contact, we measured the amount of time that the user spent with Shiverbot. This was done on the computer that also ran the voice changer software. A separate timer measured the amount of time that the participant spent cranking. Time spent cranking is a more important piece of information because some participants may simply be talking to the robot without putting physical effort into helping the robot.

In addition to recording the age and gender of the participant, one of our team members took notes on the recording sheet, which we created digitally on Google forms, and code the notes afterwards. The note taker recorded the user’s emotional state, reactions to the robot, actions taken, and overall speech patterns. We also recorded the conversation topics that we have with the user, and how they responded to each topic.






Independent & Dependent Variables

The independent variable for our study was whether or not the robot is shivering, and we measured how that will affect the amount of time that people spent assisting Shiverbot (cranking the generator), the dependent variable.







Interaction Design


Contact Initiation

“Sorry! Excuse me?”
If user gets closer:
“I wouldn’t ask unless I really needed the help, but can you please turn that crank?”


“Really sorry, my owner has been gone for a while, he said he’d be back soon.”
“Is it chilly out here or is it just me?”

Possible User Responses:

  • “Is that good?”                    - Sorry yes
  • “What is this for?”               - If you could just keep me warmer a little longer?
  • “I’m sorry I have to go”       - Really? [Contact Termination]
  • “Do I just keep cranking?”  - I’m sorry just a little bit longer please.

Contact Termination

“Thank you so much! Sorry for troubling you! What’s your name, by the way? 
User says name
“Well, it was really nice to meet you, [Name]. Take care of yourself, okay?”






Interaction Design

Burk's Pentad

  • Act - What’s the action?
    Hand-cranking a device to warm up the robot

  • Scene - What’s the context?
    Outdoors in the cold at CMU

  • Agent - Who?
    Individuals passing through CMU (e.g., students, faculty, visitors, etc.)

  • Agency - How do agents act?
    Sees shaking robot in cold, gets closer to robot, cranks robot for as long as they want

  • Purpose - Why do agents act?
    Feels compelled to see what’s wrong and help robot get warmer






Interaction Design

Tang’s Attention Commitment 








For the final design, we decided to use acrylic and sheet metal for the robot to simulate coldness of touch when people comes into contact with the robot.










In total, people spent an average of 16 seconds interacting with Shiverbot, and an average of 7 seconds using the crank.

  • Total Time Spent (Mean):                   16 s
  • Time Spent (Median):                         10 s
  • Total Time Cranking (Mean):            7 s
  • Total Time Cranking (Median):         5 s

When we broke down the data to shivering and nonshivering, we were surprised to find that both the mean time of interaction for total time spent and time spent cranking were higher when Shiverbot is shivering.

However, we also conducted two unpaired T-Tests to test for statistical significance, one for total time spent and one for total time cranking. Both of the p-values are much greater than 0.05. We did not have sufficient evidence to reject the null hypothesis that people spend an equal amount of time cranking regardless of whether or not the robot is shivering.





There were a large number of passersby that we did not count as users because they did not interact with Shiverbot on any level. This may be a result of conducting the study on CMU’s campus, as those who frequent CMU tend to be desensitized to the abundance of robots here.

Although we found that the mean for total time spent and time spent cranking are both higher for when Shiverbot is shivering, we were not able to reject the null hypothesis because of the lack of sufficient evidence. Because of various constraints around the project, we were not able to gather as much data as we had originally wanted. With more data, we can see the possibility of rejecting the null hypothesis.




Research Paper

Research Presentation

Autonomy Design Presentation

Autonomy Design Report

Team - Adella Guo, Rachel Nakamura, Alan Qiu