Janitobot fully assembled and operational in a CMU hallway, iPad mounted at neck

Human-Robot Interaction · Carnegie Mellon 2015

Helping a Robot Janitor

Does a user's perception of a service robot's level of autonomy encourage them to help it when it fails?

The Question

What kind of conditions promote ‘helping’ a robot in need and how does the perceived level of autonomy of a robot change this helping behavior? Do humans have the capacity to feel empathy towards a robot whose sole purpose is to help humans?

In order to investigate these questions, for our Human-Robot Interaction final project, we employed a service robot called Janitobot, a robot whose job was to clean the floors of Carnegie Mellon University. In our experiment, Janitobot makes verbal appeals for help to passer-bys. Across three different conditions in which Janitobot’s perceived level of autonomy was varied, we investigated how willing these passers-by were to help.

Literature Review

We grounded the experiment in prior research on trust, empathy, and perceived reliability in robotics. The central tension: humans are more willing to help other humans than machines, but robots that appear to have limited agency might trigger a different social response.

Whiteboard session with research papers and sticky notes exploring themes of trust and autonomy — Literature synthesis session — mapping the relationship between perceived autonomy, trust, and helping behavior.

Robot & Interaction Design

We settled on a custodial robot to ground the experiment in a believable, everyday scenario. Janitobot’s job was to vacuum up a spill (M&Ms and Skittles) in its target area. It would fail mid-task and make a verbal appeal for help to whoever passed by.

CAD render of Janitobot showing vacuum hose arm connected to a chest basket, with a tablet at the neck — CAD render by Yooyoung Ko. The vacuum hose connects to a chest basket; the iPad at the neck displays either a human operator, his office, or the words 'I am an autonomous robot.'

We used Burke’s Dramatic Pentad — a framework from rhetoric — to design the full staged interaction: who the robot was, what it was doing, where it was doing it, how it behaved, and why. This kept the encounter feeling real rather than theatrical.

Burke's Dramatic Pentad diagram applied to the robot interaction staging — Burke's Dramatic Pentad applied to the Janitobot encounter — ensuring the scene read as a credible, real-world event.

Experiment: Three Conditions

We ran an uncontrolled field experiment in a low-traffic area of campus, varying Janitobot’s perceived level of autonomy across three conditions. The robot made the same verbal appeal in all three — we observed whether passers-by stopped and helped.

Condition	Setup
Operator Present	Robot’s human controller clearly visible nearby
Semi-Autonomous	Controller perceived to be absent; robot appeared to act independently
Fully Autonomous	Robot appeared to act entirely on its own, no human involvement

Final assembled Janitobot in hallway with iPad neck display showing operator — Janitobot fully operational. The vacuum hose feeds into a wastebin to maintain the illusion of a functioning cleaning system.

Participants crouching to help the robot pick up spilled M&Ms — Participants helping in the Operator Present condition. Both the operator-present and semi-autonomous conditions generated substantial helping behavior.

Findings

Participants were significantly more likely to help when the robot appeared to have a human operator involved — suggesting that perceived human accountability, not robot autonomy alone, drives helping behavior. The fully autonomous condition generated the least help.

Role

Interaction designer and experimenter. Responsible for the staged interaction design, experiment protocol, and field observation.