1 of 13

1

Impacts of Robot Learning on User Attitude and Behavior

Nina Moorman

Erin Hedlund-Botti

Mariah Schrum

Manisha Natarajan

Matthew C. Gombolay

2 of 13

The world’s population is aging

2

https://thehill.com/changing-america/enrichment/arts-culture/567769-louisiana-hospitals-overwhelmed-ambulances-have-to/

3 of 13

Care robotics can help maintain independence

3

Many possible tasks

https://www.istockphoto.com/

Changing needs over time

unstructured residential environment

4 of 13

4

On-site learning for personalized care

Pre-programmed functionality

Supplemental on-site learning

5 of 13

Impact of robot learning on robot perceptions

5

Low Involvement High Involvement

Download

Reinforcement

Learning

Interactive Reinforcement

Learning

Learning from

Demonstration

Because this robot learning will likely be visible to the end-user, and can even involve the user, the way in which the robot learns may have an impact on the end-user’s perception of the robot. We investigate how the method of learning – broken into conditions representing differing degrees of user involvement in the robot learning – impacts perceptions of the robot.

LFD) In the high-involvement case, we have learning from demonstration, where a user provides kinesthetic demonstrations to the robot.

IntRL) In the medium involvement case, we have interactive RL, where the robot learns through trial and error with the environment, and this learning is supplemented with human feedback in the form of binary good or bad labels.

RL) In the low involvement case, we have reinforcement learning where the robot learns through trial and error, with no human involved.

D) Lastly, for our control condition the robot learns but the learning is not visible. Specifically, the robot acquires knowledge by downloading it from the cloud.

6 of 13

Research Questions

6

RQ1) How does robot learning method impact a user’s perception of the robot?

RQ2) Do robot perceptions differ between the general population and care givers?

RQ3) What role does physical presence play on perceptions of robots?

https://www.istockphoto.com/

Download

Reinforcement

Learning

Interactive

Reinforcement

Learning

Learning from

Demonstration

Target Population

(care givers)

General Population

(students)

Remote

In Person

We investigate three research questions in this work. First, we investigate how the method of robot learning impact a user’s perception of the robot.

Second, as our work is motivated by care robots, we include a subset of our target population, namely care givers, and we investigate whether robot perceptions differ between the general population and care giver population. Due to COVID-19, we chose to conduct the study with care givers remotely.

However, prior work has shown physical presence impacts user perception. In order to avoid confounds of care givers being remote and the general population being in person, we evaluate the general population both remotely and in person. Therefore, in our last research question, we investigate the role of robot physical presence on user attitudes and behavior with respect to the robot.

7 of 13

Study Design

7

Robot

Learning

Robot

Testing

Human Intervention

Wizard-of-Oz

General Population

60 In-person
60 Remote

Care Givers

11 Remote

Trials

7 Successful
2 Ambiguous
2 Failures

+

Participant Demographics

Our study design has two phases. First, the participant observes a video of the robot learning how to accomplish a task; this video is unique for each learning condition. Note that we employ wizard-of-oz to keep performance constant and to isolate impact of the mode of learning, independent of failure rate.

Next, in the testing phase, the participant observes the robot use what it has learned to accomplish a similar, riskier test task.

The task is plate making, where the robot must cut a banana in half with a sharp metallic knife, then place the proper medication onto the plate. We incorporate ambiguous and failure trials to ensure that we are evaluating metrics such as risk appropriately.

The participant is given the choice of intervening using an e-stop during each trial of the testing phase.

8 of 13

Metrics

8

Intervention

E-stop intervention during robot trajectory to measure reliance

Perceived Success

binary metric (handcrafted)

Perceived Safety

[Bartneck et al. 2009]

Reliability

[Körber 2018]

Demographic Information
Personality

[Donnellan et al. 2006]

Computer and Robotics Prior Experience

[Raub 1981]

Dispositional Trust

[Merritt et al. 2013]

Anthropomorphism

[Bartneck et al. 2009]

Usability and Acceptance

[Belanche et al. 2012]

Learned Trust

[Jian et al. 2000]

Task Risk

[Fischhoff et al. 1978]

Pre Study

Post Testing Trial

Post Study

Objective Metric

Quantitative Metrics

9 of 13

9

High user-involvement corresponded to lower perceived anthropomorphism

Low Involvement High Involvement

Performance appears more important than the degree of user involvement.

Impact of robot learning on robot perceptions

10 of 13

Comparing the general population to care givers

10

Care givers reported lower perceived robot success on the medicine subtask.

Loss of faith in self / confidence

Isolated medicine related risks

Loss of faith in robot

Did not intervene

Care Giver A

“Scary when it selected the wrong medication unless I read it wrong. That could be deadly.”

Care Giver B

“I should have intervened when the robot mistakenly put morphine on the plate instead of the antibiotics. I think that I would not at this time use a robot.”

Prior work has shown that caregivers feel favorably regarding robot usability and intent of future use. In our next research question, we examine how the general population’s attitudes towards learning care robots differ from that of care givers.

We find that care givers are more critical of robot success on a medicine-related task (𝑝 = .033). However, despite the lower perceived success of the robot, we do not find differences in intervention rates between care givers and the general population.

Some notable quotes obtained from our remote, caregiver participants are

“Scary when it selected the wrong medication unless I read it wrong. That could be deadly.”

Here we note that the care giver points out, and isolates the risks of improper medicine dispensing, and we see a potential low self confidence as they doubt their (correct) identification of robot failure.

In a second quote, a caregiver noted that

(2) “I should have intervened when the robot mistakenly put morphine on the plate instead of the antibiotics. I think that I would not at

this time use a robot.”

From this excerpt, we note that though they identify the error, they do not intervene. Furthermore, there is evidence of loss of trust in the care robot.

11 of 13

Investigating the impact of physical presence

11

Physically present robots are more favorably perceived.

physically present

virtual

Highlights need to consider physical presence in study design.

12 of 13

12

Include more of our target population, namely care receivers (in person)
Increase diversity of general population

Replace Wizard-of-Oz policies with learned policies

failure rate, training time, learned behavior

Place participant in the role of teacher

x

Wizard-of-Oz

Future Work

13 of 13

Takeaways

13

RQ1) How does robot learning method impact a user’s perception of the robot?

RQ2) Do robot perceptions differ between the general population and care givers?

RQ3) What role does physical presence play on robot perceptions?

Performance appears more important than user involvement.

Physically present robots are more favorably perceived.

Care givers report lower perceived success on the medicine subtask, but do not intervene more as a result.