The consumer mental model of a humanoid is a helpful butler you talk to. Factories break that model quickly. Floors are loud, workers wear hearing and eye protection and gloves, crews are often multilingual, shifts run ten or twelve hours, and every movement near a person is a safety question. Meanwhile, the people who interact with a factory robot are not one user. A line associate walks past it hundreds of times a shift and never gives it a command. A team lead gets pinged when it needs help. A remote pilot may take over its arms from another building. A plant manager only sees it as a row in an uptime report.

So the useful question for a designer is: which person, at which moment, through which surface?

Industrial humanoids moved from demo videos to paid factory work between 2025 and 2026, but the three companies I studied are taking very different bets. That difference changes what their UX looks like.

A useful reference point: Agility Robotics

Agility's Arc platform shows what mature humanoid fleet software looks like today. It covers facility mapping, workflow setup, live monitoring and troubleshooting, reports KPIs like uptime, throughput and mean time between incidents, and connects to warehouse and manufacturing systems such as WMS, MES and PLCs. The lesson: the customer-facing product is as much the fleet software as the robot.

Designing for "the user" fails here. Here are the five roles I would design for, ordered by how many of them exist on a typical floor.

Each channel has a job it is good at and a job it should never be given. The table below is my synthesis of the HRI research and factory constraints.

Three findings from the research shaped this table. First, a study at a simulated warehouse intersection found that when a robot communicated its intent, people completed tasks faster and reported more trust and comfort; a humanoid avatar using familiar warehouse hand gestures was the preferred format. Second, researchers who surveyed staff at a factory robot company found that field teams considered motion the least desirable way for a robot to show intent, and that it tended to confuse or scare people, which pushed them toward custom light signals. Third, a 2025 multimodal framework proposes matching the channel to the level of information: a flashing light for the next move, a beep for a task switch, speech for the longer-term goal.

On voice specifically, a recent systematic review of voice-based HRI found strong technical progress but limited evidence of robust deployment outside the lab. In the United States, OSHA requires a hearing conservation program when noise averages 85 decibels over a shift, a level many production areas reach, which means many workers near a humanoid will be wearing hearing protection.

Put together, an industrial humanoid needs a coordinated system of surfaces. Each layer serves a different person and escalates to the next when it runs out of bandwidth.

Every one of these companies describes its advantage as a data flywheel: deploy robots, collect real-world data, retrain, push better behavior back out. Skild names teleoperation as its richest data source. Mind Robotics calls the loop with Rivian's plants its core moat. Figure built Figure 03 to upload large amounts of fleet data for continuous learning.

The flywheel has a human bottleneck. The most valuable data comes from the cases the robot can't do yet, and those cases are resolved by people: a team lead reseating a part, a remote pilot finishing a connector mate. If that recovery is slow, confusing, or unlabeled, the line loses time and the model learns nothing.

So I think the highest-leverage UX in industrial robotics right now is exception handling. It touches every layer of the stack in about ninety seconds:

How many robots can one remote pilot supervise before time-to-assist climbs? Should the robot ever ask a nearby associate for help directly, or always route through a team lead? How should a humanoid communicate a safe power-down, given that a balancing robot can't simply be switched off the way a fixed arm can? How should light and voice behave when two robots from different vendors share an aisle? And which failure causes do line workers actually see most often, versus what the logs report?

Safety standards will shape many of these answers. ISO 10218 was revised in 2025, and ISO 25785-1, the first standard written for industrial mobile robots that must actively balance, is still in committee draft. The interaction patterns designed now will need to fit whatever those standards require.

Method: desk research across company announcements, job postings, trade press and published HRI research, current as of September 2026. The inferences marked above are my own and have not been confirmed by the companies. My next step is interviewing people who work on factory lines to test the stack and the assist flow.