In a warehouse pilot earlier this year, a humanoid robot executing a bin-retrieval task paused mid-route to recalculate a path around an unexpected obstruction. To the human worker approaching from the opposite direction, the pause looked like a malfunction. The worker stopped, waited, then walked a wide arc around the robot rather than risk getting in its way. The robot eventually resumed its task. No incident occurred. But the interaction left the worker unsettled enough to mention it to a supervisor — not because the robot had done anything wrong, but because they had no idea what it was doing or why.
That anecdote, drawn from published field research on human-robot collaboration in logistics settings, captures a challenge that gets relatively little attention in the public coverage of humanoid robotics: the communication problem. Building a robot that can physically navigate a human environment is hard. Building one that humans can actually work alongside — comfortably, safely, and productively — requires something additional. It requires the robot to be legible.
What Legibility Means in a Workplace
When two people share a physical workspace, they communicate constantly without speaking. A glance communicates that someone is about to step into your path. A shift in body weight signals that a person is about to change direction. The way someone reaches toward a shared object communicates whether they're taking it or just examining it. These signals are so automatic that we rarely notice them — until they're absent.
Human-robot interaction researchers use the term "legibility" to describe a robot's ability to make its intentions readable to nearby people. A legible robot doesn't just execute a task correctly; it executes the task in a way that communicates what it's about to do before it does it. This turns out to be a genuinely difficult design challenge, and it's distinct from the mechanical question of whether the robot can perform the task at all.
The research on this is reasonably well-established in the academic literature. Studies going back more than a decade have documented that humans working near robots feel significantly more comfortable — and make fewer errors — when the robot's movements are predictable and communicative, even when the robot's actual task performance is identical. The same motion, made in a more telegraphed and deliberate way, is experienced as safer and less stressful than an equivalent motion that is mechanically correct but abrupt or hard to read.
For early-generation humanoid robots, which tend to move in ways that are neither fully human-like nor consistent with any other recognisable motion pattern people encounter in daily life, this is a live problem.
The Specific Signals Humans Use — and Robots Mostly Don't
Consider what humans do when approaching a shared workspace. We slow down before we arrive. We orient our bodies toward where we're going, which gives bystanders early information about our trajectory. We pause and make eye contact — or at least establish mutual awareness — before reaching into a shared space. We use brief sounds or words to mark intent: a soft "excuse me," a clearing of the throat, sometimes just a change in our breathing pattern that signals we're about to do something.
Current humanoid robots replicate very few of these signals reliably. Most don't have functional eye contact — their cameras capture their environment but don't communicate attention in a way humans can read. Their approach patterns are often optimised for path efficiency rather than communicative value. Their vocalisation capabilities, where they exist at all, are either absent or limited to explicit verbal statements that feel jarring in informal workplace contexts.
Some companies are working directly on this. Boston Dynamics' Atlas has demonstrated expressive motion — movements that go beyond strict task efficiency to signal something about the robot's state or intent. Sanctuary AI's Phoenix robot uses a combination of posture and pacing changes to signal task transitions. Figure's robots have demonstrated basic acknowledgement gestures in demo settings. But whether these capabilities hold up in the variety and unpredictability of real workplaces, with real workers who haven't been briefed on how to read them, is a different question from whether they work in a controlled demonstration.
The Asymmetry of Adaptation
There's an implicit assumption in much of the humanoid deployment literature that the communication gap will be bridged primarily by humans adapting to robots. Workers will be trained on how to interact with robots. New protocols will be established. People will get used to it.
This assumption is worth scrutinising. The history of industrial automation does show that workers adapt — often substantially — to new machinery. People who work daily with robotic arms, conveyor systems, and autonomous forklifts develop strong intuitions about how those systems behave. Adaptation is real.
But there are limits. Adaptation to predictable, domain-specific machinery is easier than adaptation to a general-purpose humanoid robot that moves differently from any other machine in the environment, whose range of tasks is variable, and whose decision-making process is not transparent. The cognitive load of maintaining awareness of a humanoid robot's intentions — when it might move, where it's going, whether it has seen you — is meaningfully higher than the load associated with a conventional fixed-path automated system. That load has real costs: stress, distraction, and in some documented cases, reduced productivity in the human workers sharing the space.
The research on this point is still developing, and most existing studies involve relatively short exposure periods in controlled settings. What happens to human-robot communication dynamics over weeks and months of continuous co-working is less well understood.
Voice, Screens, and Light — the Current Toolkit
The approaches currently being developed to bridge the legibility gap fall into a few categories. Voice and sound: some humanoid platforms can emit tones, chimes, or verbal signals to mark approaching movement or task transitions. Visual displays: a number of designs incorporate small screens on the robot's torso or head that can display simple status information — a directional arrow, a task icon, a colour-coded state indicator. LED lighting: edge-lit panels or indicator lights around a robot's body can signal direction of travel, proximity warnings, or operational state in a format that's visible peripherally without requiring direct eye contact.
Each of these has practical limitations. Voice signals are useful in relatively quiet environments but are easily masked in noisy warehouses or manufacturing floors. Screens require workers to look directly at the robot to read them — the opposite of the peripheral-awareness model that human-human communication relies on. LED patterns need to be learned before they're meaningful, which means they work well for trained workers but provide no information to someone encountering the robot for the first time.
None of these solutions approach the richness or naturalness of human nonverbal communication. That gap may be acceptable for highly structured, purpose-built deployment environments where workers are trained and the robot's task range is limited. It's a more significant obstacle for the general-purpose humanoid use case that most companies are eventually aiming for.
What "Human Form Factor" Was Supposed to Solve
Part of the original argument for humanoid form — two arms, two legs, operating at human scale — was that it would make robots more intuitive to work alongside. A robot that looks roughly human would, the reasoning went, be easier for people to read and predict than an industrial arm or a wheeled mobile platform.
The evidence on this is mixed. The human form does give people a mental model to start from — we know roughly what a biped can and can't do, which is more than we know about a novel mechanical platform. But humanoid robots that are imperfect approximations of human movement can trigger what researchers sometimes describe as an "uncanny valley" effect in motion: movements that are almost human but not quite generate more discomfort than movements that are clearly non-human. A robot that walks almost like a person but not quite, that turns its head almost like a person but not quite, produces a subtly unsettling effect that a purely mechanical system does not.
Whether current humanoid platforms have crossed this threshold in workplace settings is genuinely unclear. The pilot data from active deployments hasn't been published in a form that allows independent assessment. What can be said is that the humanoid form factor doesn't automatically solve the legibility problem — it reframes it, in some ways makes it harder, and requires deliberate design attention to address.
A Problem That Gets Harder at Scale
The communication challenge is manageable in a deployment of five or ten robots in a carefully prepared environment with a trained workforce. It becomes more complex at the scale that humanoid companies are projecting. When multiple humanoid robots are operating in the same space, when workers haven't been specifically trained on each robot's behavioural patterns, and when the task repertoire expands beyond the narrow scope of current pilots, the legibility problem multiplies.
This isn't an argument against the technology. It's an argument for taking the communication design as seriously as the mechanical design. The companies that figure out how to make their robots genuinely easy to read and work alongside — not just mechanically capable — will have a meaningful advantage in the operational contexts that matter most. That design challenge is underappreciated in the current coverage, and it deserves more attention than it's getting.