articles · studio notes
Why operators trust some robots and route around others.
On a working floor, trust isn't a feeling. It's a behaviour: whether people work near the machine or route around it. As more robots leave the fenced cell for floors they share with people, and start taking work from scheduling software and plain-language requests as well as a pendant, that behaviour matters more. It's still designed, long before the machine ships.
skeelx · updated oct 2026 · 4 min read
Trust is prediction
People trust what they can predict. A robot earns floor trust the way a good colleague does: it telegraphs what it's about to do, it does the same thing the same way every time, and when something's wrong it says so before anyone has to guess. Most "workers won't accept the robot" problems are really prediction problems: a machine whose intentions are illegible from three metres away, moving at speeds that read as sudden even when they're safe. The more flexibly a robot can be told what to do, the more its motion has to explain itself: an operator who didn't teach today's task can't predict it from memory.
The design levers
Legibility of motion: approach paths and accelerations that read as deliberate; pauses before direction changes; speed profiles tuned to how they look, not just what the safety case permits. Honest status: state made visible at a glance (working, waiting, degraded, stopped-and-why), with light and sound designed as a language rather than decoration. The worst state signal is silence that could mean anything. Form that signals intent: silhouette, guarding and detailing that tell the floor what this machine is for and where its reach ends. Recovery the operator owns: when it faults, the path back should be legible to the person standing next to it, which is the entire argument of designing the fault state first.
The pendant is the relationship
Operators meet the robot through its interface far more than through its payload. A teach pendant or console that respects the operator (plain states, forgiving flows, no punishment for curiosity) builds the trust that spec sheets can't. This is HMI work under real constraints, and it's where "acceptance" is mostly won or lost. As more ways in appear (a tablet, a spoken request, a task sent by fleet software), the pendant's job grows rather than shrinks: it's where the operator sees what the machine is doing next, and who asked for it.
The agent-native lens: one state language for people and schedulers
When the buyer sends an assistant. An operations team shortlisting robots may hand the first pass to an assistant, and it weighs what is written down. Spec sheets are thick with payload and reach and thin on how the machine behaves near people. Publish the trust levers as plain facts: the states the robot shows and how each one looks and sounds, what an operator can recover without calling an engineer, how it signals a change of direction. On site, its status gains a second reader: the fleet-management software and scheduling agents that hand it work. Give the light, the pendant and the API one state vocabulary, so "waiting" means the same thing to the person at the cell and to the software dispatching the next task.
When the business runs on agents. Distrust tends to show up in machine data before it reaches a budget review: repeated protective stops in one zone, a cell left at reduced speed for whole shifts, manual overrides clustered around the same task. Agents can watch a fleet for those patterns, group them by site and task, and draft a review for the team that tunes motion and interface. Where an agent assigns the robot's work, the person at the cell can still pause it. Changes to speeds, zones or safety settings are never pushed by an agent; an engineer responsible for the safety case approves them, and the site signs off before they run.
Trust as commercial strategy
Trusted machines get used, which means they generate the uptime and the stories that sell the next ones. Distrusted machines get bypassed, and bypassed machines get returned in the next budget cycle. For robotics companies, HRI isn't a polish layer. It's retention engineering, and it gets harder to retrofit as robots take on more varied work beside people and more of their instructions come from software. That's why our robotics sector work treats the operator's first shift with the machine as a design deliverable in its own right: whoever, or whatever, sends the task, the person standing next to the machine should never have to guess what it will do.