Campus Human–Robot Interaction
Robotic Receptionist
A campus receptionist prototype that explores how people interact with a robot when asking for directions or student services.
The problem
Visitors to a university lobby ask the same questions again and again: where an office is, which floor a department is on, who can help with a student service. Staff are not always at the desk.
Our graduate team is prototyping a lobby receptionist for Bay Atlantic University that can greet visitors, answer common questions, give directions, identify the right department, and hand off to staff when it should.
The project sits between AI and service design. Most of the hard questions are about when the robot should answer, when it should ask, and when it should step back.
Constraints
- Facts such as office locations must always be correct, so they cannot depend on a model’s guess.
- Visitors should be able to reach a human whenever the robot is unsure.
- Course scope: prototype demos and requirements, not a permanent installation.
My role
- Own the computer vision and human–robot interaction area of the project.
- Wrote the vision-related functional requirements (FR-V) and the interaction scenarios used in demos.
Teammates lead testing, UX, and documentation and data. Faculty and a mentor advise the project.
System design
The interaction loop separates what the robot notices, what it decides, and what it says, so each part can be tested on its own.
- NoticeVision detects that a visitor is approaching and ready to interact.
- UnderstandThe visitor’s question is mapped to an intent such as directions or a service request.
- DecideFixed rules answer factual questions; uncertain cases go to clarification.
- Respond or escalateThe robot answers, asks a follow-up, or hands off to staff.
Tradeoffs
Fixed rules for facts, learning for intent
Why Office locations and procedures are stored as fixed rules. Only intent recognition is learned, so a model error leads to a clarifying question rather than a wrong room number.
Cost Every campus change has to be updated in the rules by hand.
What I can show
Related course experiment
In a separate Applied AI lab, I trained a small decision tree intent classifier on synthetic receptionist questions, tuned max_depth, read the learned rules, and compared it with an SVM. It is a classroom experiment, not the team prototype.
Requirements and scenarios
My vision requirements and interaction scenarios guide the prototype demos.
What I would explore next
- Measure how often visitors need escalation in demo sessions.
- Test how wording affects whether people trust the robot’s directions.