Human–Robot Interaction
What Is Human–Robot Interaction?
The field that studies robots and people together: its three levels of interaction, the roles a human can play and why it borrows as much from psychology as it does from engineering.
beginner · 13 min read
Lesson content
Every path before this one treats the robot as a system to be built.
This one treats the robot and the person as a single system to be designed together.
The reason is simple. A robot that works flawlessly and that nobody can work alongside has not solved the problem it was bought to solve.
Definition
Human–Robot Interaction (HRI)
The study of how people and robots work together, and the design of robots that can deal with people well: physically, socially and cognitively.
Formal definition
A multidisciplinary field concerned with understanding, designing and evaluating robotic systems for use by or with humans, drawing on robotics, human–computer interaction, psychology, cognitive science, design and ethics.
Is this not just interface design for robots?
It overlaps a great deal. Three things make it genuinely different.
A robot takes up space. It can block you, bump you or stand too close. None of that is possible for a screen.
A robot acts on its own. It does things you never asked for, which means it has to be predictable in a way an app never needs to be.
A robot gets read as an agent. People hand intention, personality and even feeling to anything that moves under its own power.
That last one is worth sitting with. It is not a design choice you can opt out of. It is a fact about human perception, and every HRI decision gets made in its shadow.
- Physical interaction
The robot and the person share space and sometimes contact. Handing over a tool, guiding an arm, working at the same bench.
Safety dominates everything else. Forces, speeds and stopping distances have to be right before any other question is worth asking.
- Social interaction
The two exchange signals: speech, gaze, gesture, expression and turn-taking.
This decides whether working with the robot feels natural or exhausting. Being able to tell what it is about to do matters more than how clever it is.
- Cognitive interaction
The two share a task, a plan or a picture of the world: who does what, what happens next, what went wrong.
Shared mental models and mixed-initiative teamwork live here. This is what turns a tool into a colleague.
The human is not always in the same role. The role changes what the robot has to be good at.
Five are usually distinguished. The difference between them comes down to two things: how much training that person has had, and how much say they had in the robot turning up at all.
| Role | What the person does | What the robot owes them |
|---|---|---|
| Supervisor | Watches a fleet and steps in rarely | Clear status, good alerts, a graceful way to degrade |
| Operator | Commands the robot directly | Responsive control, low latency, honest feedback |
| Teammate | Works beside it on a shared task | Readable intent, predictable motion, decent turn-taking |
| Bystander | Shares the space without commanding anything | Safe behaviour that assumes no training whatsoever |
| Mechanic or programmer | Configures it, teaches it, repairs it | Diagnostics, teach interfaces, honesty about its own state |
Supervisor
- What the person does
- Watches a fleet and steps in rarely
- What the robot owes them
- Clear status, good alerts, a graceful way to degrade
Operator
- What the person does
- Commands the robot directly
- What the robot owes them
- Responsive control, low latency, honest feedback
Teammate
- What the person does
- Works beside it on a shared task
- What the robot owes them
- Readable intent, predictable motion, decent turn-taking
Bystander
- What the person does
- Shares the space without commanding anything
- What the robot owes them
- Safe behaviour that assumes no training whatsoever
Mechanic or programmer
- What the person does
- Configures it, teaches it, repairs it
- What the robot owes them
- Diagnostics, teach interfaces, honesty about its own state
Example
One robot, four roles at once
A hospital delivery robot has four different humans around it before lunchtime.
A supervisor in the logistics office watching a dashboard.
An operator, a nurse who loads it and sends it somewhere.
Bystanders: patients and visitors in the corridor who never agreed to any of this.
A mechanic who swaps its batteries on Thursdays.
Every one of them needs something different from the same machine.
The hardest of the four is the bystander. No training, no briefing, no say in whether a robot turned up in their day. Designing only for the operator is the most common HRI failure there is.
Autonomy is not a switch. It runs along a spectrum, and where a system sits on that spectrum decides how much of the interaction burden lands on the person.
Levels of autonomy, and what each one asks of the human
1 · Teleoperation
The human does everything and the robot is an extension of their body. It needs excellent feedback, low latency and enough of the robot's sensing to know what is going on out there.
2 · Assisted teleoperation
The robot helps: collision avoidance, motion scaling, tremor filtering. The human still leads but the robot can veto a bad command. Surgical robots live here.
3 · Shared control
Both contribute all the time. The human sets the intent and the robot handles the execution. Hand-guiding an arm is exactly this.
4 · Supervised autonomy
The robot runs the task while the human watches and steps in. The hard part is keeping that human awake enough to step in well, because attention decays fast when nothing happens.
5 · Full autonomy
The robot handles everything inside its envelope. The interaction moves entirely to before, when the task is specified, and after, when someone works out what it actually did.
The irony of automation
Here is an uncomfortable pattern. The better a robot gets, the worse the human becomes at supervising it.
Follow the chain. Interventions get rare, so the person gets no practice. Reliability is high, so vigilance drops. Then the moment finally arrives, and an out-of-practice person has to diagnose an unfamiliar situation in a few seconds.
Aviation documented all of this decades ago.
Which is why "add more autonomy" is not automatically the safer choice. Autonomy has to be designed around the handover, not just around the steady state.
HRI gets measured, not asserted.
Because the outcome lives partly inside somebody's head, the field borrows its methods from psychology: controlled studies with real participants, questionnaires that have been validated, and behaviour rather than opinion wherever behaviour is available.
- Godspeed questionnaire
A standard set of scales covering anthropomorphism, animacy, likeability, perceived intelligence and perceived safety.
The most widely used instrument in HRI, so a result that uses it can be compared against other people's results.
- Wizard of Oz study
A hidden human quietly drives the robot while participants believe it is running itself.
It lets you test an interaction design before the autonomy exists. Standard practice. It has to be disclosed when you debrief people.
- Task performance measures
Completion time, error rate, number of interventions, workload.
Objective. They often disagree with what participants said they preferred. Measure both and look at the gap.
- Situation awareness
How well the human knows what the robot is doing, why it is doing it and what it will do next.
The single strongest predictor of whether a human and a robot perform well as a team.
Designing the robot first and the interaction afterwards.
By then the shape, the speed and the sensing have already decided what the interaction can be. Interaction requirements belong in the concept phase.
Testing only with the engineers who built it.
They know what every light means and where the stop button is. Test with people who look like the real operators and the real bystanders, including the ones who did not choose to be there.
Treating "people liked it" as a result.
Novelty makes almost any robot likeable for ten minutes. Measure across repeated sessions if you want to know whether it will still be switched on next month.
Key takeaway
- HRI designs the robot and the person as one system rather than designing the robot alone.
- Interaction runs on three levels at once: physical for safety, social for signals, cognitive for the shared task.
- The person may be a supervisor, an operator, a teammate, a bystander or a mechanic. The bystander is the hardest.
- Autonomy is a spectrum, and more of it moves the interaction problem rather than removing it.
- HRI claims get tested with human-subject methods, because the outcome lives partly inside somebody's head.
Finished this lesson?
Guest progress is saved on this device.