Count the humanoid robot pilots announced over the past two years and you get an impressive number. Amazon and Agility Robotics. BMW and Figure AI. Foxconn evaluating multiple platforms. Logistics companies across the United States, Germany, and Japan running trials with hardware from at least five different manufacturers. The announcement flow has been consistent enough that a casual reader of the sector might reasonably conclude that humanoid robots are steadily moving from labs into operations at scale.
That conclusion would be premature. What is actually happening — and what the announcements rarely capture — is that the humanoid robotics industry has developed a significant structural problem: pilots start, pilots run, and pilots frequently do not expand into the sustained, scaled deployments that the investment thesis for the entire category depends on. The gap between a functioning pilot and a commercial deployment is proving considerably wider than either the companies or their customers initially expected.
Understanding why that gap exists, and what it would take to close it, is more useful than tracking which company announced a trial with which customer this month.
What a Pilot Actually Involves
When a humanoid robotics company announces a deployment partnership with a major industrial customer, the operational reality is typically more constrained than the language implies. A pilot usually involves a small number of units — often between two and ten robots — operating in a designated area of a facility, executing a specific and deliberately narrow set of tasks, with dedicated engineering support from the robotics company on-site or on call. The customer provides the physical space and cooperates on task selection. The robotics company provides the hardware, handles maintenance, and often stations engineers on site for the duration.
This structure is sensible for both parties. The robotics company gets access to a real industrial environment and real operational data without having to commit to commercial terms it cannot yet sustain. The customer gets to evaluate the technology without significant capital outlay or operational risk. The costs of failure are bounded on both sides.
The problem is that this structure also creates conditions that are systematically more favourable to the robot than a genuine commercial deployment would be. The task is chosen to suit the robot's current capabilities. The engineering support is immediate and expert. The environment is adjusted — sometimes subtly, sometimes substantially — to reduce friction for the hardware. When the pilot produces positive results, those results are real, but they reflect performance under conditions that will not exist when the robot is deployed at scale without dedicated support and without task curation.
Where Pilots Break Down
The transition from a supported pilot to commercial deployment typically surfaces several categories of problem that the pilot structure obscured.
The first is task variability. In a well-designed pilot, the robot's task envelope is kept narrow enough that the system can handle it reliably. In production, the same facility has more variability: packaging changes, floor layouts shift, ambient conditions differ across seasons, the mix of objects the robot encounters expands beyond what the pilot covered. Each additional variable is, in principle, manageable. In aggregate, they compound in ways that current humanoid systems handle unevenly.
The second is the support model. A pilot with two engineers from the robotics company on-site has a fundamentally different reliability profile than a fleet of forty robots supported by a standard service contract. The skills required to maintain a humanoid robot — diagnosing sensor failures, recalibrating manipulators, troubleshooting software edge cases — are not yet widely available in the industrial maintenance workforce. The skill gap that the industry is facing (a subject covered separately here) means that scaling a deployment requires either a large and expensive service organisation from the robotics company or a significant investment in customer-side technical capability. Neither option is cost-free, and neither is fully accounted for in the headline economics of most pilot announcements.
The third is integration complexity. A pilot can run alongside existing systems without deep integration. A scaled deployment, by contrast, typically needs to connect with warehouse management software, safety monitoring systems, fleet coordination platforms, and shift management processes. That integration work is time-consuming, requires cooperation from multiple internal teams at the customer, and often surfaces compatibility problems that the pilot never encountered because it was designed to avoid them.
The Customer's Calculation
From the customer's perspective, the decision to expand a pilot into a scaled deployment involves a cost-benefit calculation that the pilot itself only partially answers.
The pilot answers the question: can this robot execute this task in this environment? That is a necessary condition for scaling, but not a sufficient one. The customer also needs answers to questions the pilot is rarely designed to address: What is the true cost per task at scale, including maintenance, downtime, and integration overhead? How does the robot's performance hold up over twelve or twenty-four months of continuous operation? What happens to throughput when one robot in a fleet of twenty fails? What is the contractual recourse if the robotics company's service response falls short?
These are the questions that procurement and operations teams raise when a pilot transitions from the innovation team to the people who will actually operate the technology at scale. They are also the questions that humanoid robotics companies are least equipped to answer with data right now, because their commercial deployments are too recent and too limited to have generated the operational track record that answers them.
The result is a predictable dynamic: pilots produce genuine technical validation, the customer's innovation team is enthusiastic, and then the proposal to scale meets friction from finance, operations, and procurement teams who are applying standard criteria for capital equipment investment to a category that cannot yet meet those criteria. The pilot does not fail. It simply does not convert.
The Announcement Incentive
One consequence of the pilot problem is that it creates a structural incentive for humanoid robotics companies to generate a continuous stream of pilot announcements without public accountability for what happened to the previous round. A new pilot partnership with a recognisable industrial customer is press coverage, investor confidence, and demonstrated market traction. A pilot that concluded without expanding into a commercial deployment is a data point that rarely appears in public communications.
This asymmetry is not unique to humanoid robotics — it is a feature of most enterprise technology sales cycles. But it is particularly pronounced in humanoid robotics because the investment stakes are high, the timelines are long, and the technical complexity of the subject makes independent verification of commercial progress exceptionally difficult. There is no external source that systematically tracks how many humanoid robot pilots have been announced, how many have concluded, and how many have converted to commercial deployments. The information is entirely within the control of the companies involved.
What this means for anyone trying to assess the industry honestly is that announcement volume is a poor proxy for commercial progress. The relevant question is not how many pilots are running, but what fraction of completed pilots result in expanded commercial commitments — and that number is not publicly available for any major player in the sector.
What Closing the Gap Requires
The pilot-to-scale gap is not a permanent feature of the industry. It is a stage in the maturation of a technology category, and several things could close it.
The most direct route is reliability at scale over time. If early commercial deployments — the ones that have converted from pilots — demonstrate high uptime, predictable maintenance costs, and task performance that holds up over months rather than weeks, those results become the evidence base that operations and finance teams need to approve expansion. Every successful scaled deployment makes the next one easier to approve. The industry needs those reference deployments to exist and to be documentable in terms that procurement teams can evaluate.
The second route is cost reduction. Many of the economics problems that prevent pilots from converting are directly related to the current cost of hardware and support. As unit costs decline and service infrastructure matures, the cost-per-task calculation shifts, and deployments that were marginal at current costs become clearly viable. This is the standard hardware category maturation dynamic, and it is plausible — but it requires either time or capital deployment at a scale that only a few companies in the current landscape can sustain.
The third route is task standardisation. One reason industrial robots in fixed configurations achieved widespread deployment faster than humanoid robots is that the task set was defined narrowly enough that the technology could be optimised for it completely. Humanoid robots' selling point — adaptability, general-purpose utility, the ability to operate in environments designed for humans — is also what makes them harder to deploy reliably. Companies that identify specific high-value, high-repetition tasks and treat them as product-level commitments, rather than demonstrations of general capability, are likely to convert pilots faster than those leading with flexibility.
None of these routes is fast. The pilot problem will not resolve in a single product generation, and the industry's current announcement cadence suggests that not everyone building humanoid robots has fully internalised that the gap between impressive pilot and scaled deployment is where the commercial case actually gets made or lost. The robots that matter, ultimately, are not the ones in the pilot. They are the ones still running two years later, without the engineering team from the manufacturer on-site, handling tasks the original pilot designers did not anticipate. That is the bar, and the industry is still some distance from clearing it routinely.