Human-Robot Interaction, Explained for the People Actually Sharing a Floor With One
Human-robot interaction (HRI) sounds like it belongs in a conference paper, and it does—HRI is a real academic field, with its own conference (ACM/IEEE HRI), its own journals, and decades of research on trust, communication, and collaboration between people and machines. But strip away the terminology, and it's also the most concrete question a plant manager asks before a deployment: what actually happens when a person and a robot are in the same aisle, at the same time, doing related work?
The honest answer has less to do with the robot's intelligence than most vendors imply. It has to do with a specific set of engineering choices—how the robot signals what it's about to do, how much a worker trusts it, and how much of the interaction is actually designed rather than assumed. We've written before about what humanoid robots actually do in deployment and why the robot itself is the easy part of the ROI equation. Interaction design sits inside both of those problems. It's not a UX nicety layered on afterward; it's a safety standard, a trust variable, and a line item in your deployment budget, whether or not anyone wrote it down as one.
The part that's a standard
Before "human-robot interaction" was a design philosophy, it was a regulatory requirement. Collaborative robots working near people without a safety cage are governed by ISO 10218 and its companion specification, ISO/TS 15066, which define four distinct modes of interaction a robot is legally allowed to have with a nearby worker [1]:
Safety-rated monitored stop—the robot halts completely the moment a person enters its workspace and won't resume until the space is clear.
Hand guiding—a worker physically guides the robot's motion, typically via a control button on the robot itself.
Speed and separation monitoring—the robot continuously tracks the distance to the nearest worker and slows or stops as that distance closes, using laser scanners, radar, or 3D vision [2].
Power and force limiting—the robot's joints are engineered so that even unplanned contact stays below a threshold that would cause injury, calibrated per body region [2].
Each of these is a different interaction design, not just a safety feature bolted onto one underlying robot. A robot built around speed and separation monitoring behaves completely differently around a person than one built around power and force limiting—one avoids you, the other is built to survive touching you. Picking the wrong one for the task isn't a compliance footnote; it's the difference between a robot a worker can ignore and one that stops their line every time they walk past it.
Trust is the variable nobody puts on the spec sheet
Get the safety engineering right and there's still a harder problem: whether the person working next to the robot trusts it appropriately. A 2025 systematic review of 58 studies on trust in human-robot collaboration found that trust in these settings behaves asymmetrically: it builds slowly, through a robot demonstrating reliability and transparent behavior over time, but can collapse suddenly after a single failure or unexpected action [3]. The same review is explicit that both directions of miscalibration carry real costs—overtrust produces complacency and reduced monitoring, while undertrust produces disuse and workers routing around the robot entirely, which quietly erases whatever productivity gain justified the deployment [3].
Overtrust is the less intuitive failure and the one worth taking seriously. In a widely cited study, researchers had participants follow a robot during a simulated building evacuation. When the robot visibly malfunctioned, moving erratically, even leading people toward a dead end—most participants followed it anyway, deferring to a machine they had no real basis to trust that much [4]. That's not a robot problem. It's a design failure in the interaction: nothing about the robot's behavior signaled "I am currently wrong," so nobody course-corrected. On a factory floor, the equivalent failure looks less dramatic but is just as costly—a worker who stops watching a robot's payload because it's "always been fine," right up until the one time it isn't. This is exactly the kind of accountability gap we wrote about in our piece on robot ethics on the factory floor: someone specific has to own the decision to intervene when a robot is behaving outside its tested envelope, and trust calibration is the mechanism that makes that intervention actually happen instead of getting waved off.
The interaction depends entirely on what kind of robot is in the room
Not every humanoid is designed to interact with people the same way, and conflating them is where a lot of buying decisions go wrong. In our breakdown of the fourteen real categories of humanoid robots, the difference is stark once you look for it. A robot like Engineered Arts' Ameca is built almost entirely around interaction as the product—dozens of facial expressions, conversational AI integration, and a design goal of holding a person's attention and building rapport in a lobby or a museum. A robot like Agility's Digit or Figure's Figure 03 is built to be worked around, not with—it's interaction budget goes almost entirely into predictable, legible movement in a shared aisle, not conversation. Buying an industrial humanoid and evaluating it on how personable it feels, or buying a social robot and evaluating it on payload and throughput, both miss what the robot was actually engineered to do.
This matters operationally because interaction design isn't free, and every dollar spent on one kind buys less of the other. A robot with an expressive face and natural-language conversation is spending the compute and engineering budget on a channel a warehouse aisle mostly doesn't need. A robot optimized purely for speed and separation monitoring will feel cold and unreadable in a role that actually calls for reassurance—reception, patient information, retail. Matching the interaction design to the job is a large part of why some deployments feel effortless and others generate constant low-grade friction that never shows up on a spec sheet.
Interaction is also how the robot learns — which is where consent comes back in
There's a newer form of human-robot interaction that doesn't involve a robot in the room at all: a person doing their job on camera, while a model learns from watching. World action models can now learn physics and task structure directly from video of a person performing the work, which means your existing process footage is itself training data, and the more structured version of that idea, purpose-built automated data collection is becoming the actual bottleneck in how fast physical AI improves.
That's a form of interaction too, even though nobody involved would describe it that way in the moment. A worker being filmed to train a robot's replacement task is interacting with that robot, just on a delay. Which is exactly why we've argued that the ethical bar here is low but non-negotiable: people should know, before filming starts, that footage may train a model and roughly what it will be used for. Good interaction design and good consent design turn out to be the same discipline, applied at different points in the robot's lifecycle—one governs how a robot behaves around a person today, the other governs how a person's past work behavior shapes a robot's future behavior.
Language is becoming the interface — including for the humans configuring the robot
The most underrated human-robot interaction on a factory floor isn't between a worker and a deployed robot at all—it's between an engineer and the system that configures one. World action models have made natural-language instruction a viable control interface, not just a chat gimmick, because a model that can adapt to new hardware from a short demonstration can also be steered by a plain-language description of the task rather than a hand-coded motion plan. That's the interaction model behind Bellwether: a person describes, in normal text, how they want the work done, and that description becomes the storyboard the deployment is built from. It's a form of HRI most of the academic literature doesn't cover yet, because it's newer than the field's core research—but it's arguably the interaction that determines everything downstream, since a poorly specified instruction produces a robot that behaves unpredictably around the very people the safety engineering above is designed to protect.
Why this is a deployment cost, not just a comfort issue
It's tempting to file all of this under "nice to have" relative to throughput and uptime. The data on manufacturing talent says otherwise. BCG's research on the deployment gap found that 87 percent of manufacturers identify talent — not technology — as the binding constraint on sustaining an automation program [5], and we've written before about why that statistic matters more than the productivity headline it was buried under. Poor interaction design is a direct tax on exactly that scarce resource. A robot whose behavior is illegible generates more supervision overhead, not less—workers who don't trust it hover; workers who overtrust it stop watching at all; either way, someone is spending attention on the robot that the deployment was supposed to free up. Interaction design that gets trust roughly right is one of the cheaper ways to protect the very staffing capacity manufacturers are already short on.
Frequently asked questions
What is human-robot interaction, in plain terms? It's the discipline of designing how a robot communicates its intent, state, and next action to the people around it—through movement, signals, sound, or language—so that people can work near it safely and predict what it will do next.
Is human-robot interaction the same thing as cobot safety? Related, but not identical. Cobot safety standards like ISO/TS 15066 define the physical and behavioral rules a robot must follow around people [1][2]. HRI is the broader discipline that also covers trust, communication, and how interaction design changes across robot types—safety is the floor, not the whole field.
What is "overtrust" and why does it matter on a factory floor? Overtrust is when a person relies on a robot more than its actual reliability justifies, which reduces monitoring and increases risk when the robot eventually fails or behaves unexpectedly [3][4]. It's a design failure to solve for, not a worker error to blame.
Does every robot need to be designed for social interaction? No—and trying to make every robot sociable is usually a mistake. Industrial humanoids doing warehouse or assembly work are generally better served by legible, predictable movement than by conversational features; social and service robots are the opposite case. Robot type should drive interaction design, not the other way around.
How does 3GEN think about human-robot interaction in deployment? As a design constraint that belongs in the deployment plan from day one, not a feature added after a pilot succeeds—covering everything from ISO-compliant physical safety behavior to plain-language task specification through Bellwether. Reach out if you're evaluating what that looks like for your floor.
Notes: This piece reflects editorial judgment on an active research area, not a compliance or legal reference — ISO/TS 15066 and related standards should be reviewed directly, and with a qualified safety engineer, before any deployment decision. Academic citations are drawn from peer-reviewed and industry-technical sources, not 3GEN internal research.
References
[1] Standard Bots. "Collaborative Robot Safety Standards You Must Know." 2026. https://standardbots.com/blog/collaborative-robot-safety-standards
[2] Orbbec. "Depth Cameras for Cobot Safety: ISO/TS 15066 Guide." 2026. https://www.orbbec.com/blog/depth-cameras-cobot-safety-iso-15066/
[3] Frontiers in Organizational Psychology. "How to Achieve Human-Centered Automation: The Importance of Trust for Safety-Critical Behavior and Intention to Use in Human-Robot Collaboration." October 8, 2025. https://www.frontiersin.org/journals/organizational-psychology/articles/10.3389/forgp.2025.1669782/full
[4] Robinette, P., Li, W., Allen, R., Howard, A. M., Wagner, A. R. "Overtrust of Robots in Emergency Evacuation Scenarios." Proceedings of the 11th ACM/IEEE International Conference on Human-Robot Interaction, 2016. Cited via Springer, International Journal of Social Robotics. https://link.springer.com/article/10.1007/s12369-019-00596-x
[5] 3GEN Robotics, citing BCG. "BCG Says Deployment Now Decides Competitiveness." August 2026. https://www.3genrobotics.com/blogs/bcg-factory-of-the-future-deployment-gap
Related reading: What Humanoid Robots Actually Do in 2026 · 14 Types of Humanoid Robots · Robot Ethics on the Factory Floor · Your Factory Already Has Robot Training Data