Every demonstration video of a humanoid robot ends the same way: the robot completes its task, the camera cuts, and everything looks clean and controlled. What you never see is what happens three weeks into deployment when an actuator starts behaving erratically, when a cable harness develops intermittent contact, or when a knee joint accumulates enough wear to fall outside the tolerance window for safe operation. You never see who shows up to fix it, how long it takes them, or what it costs.
This is not an incidental gap in the coverage. Maintenance and repair is one of the least-examined constraints on whether humanoid robotics can scale from small supervised pilots to the kind of widespread deployment the industry is aiming for — and it is genuinely unsolved in ways that matter for anyone trying to evaluate the realistic pace of that scaling.
The Mechanical Reality
Humanoid robots are mechanically complex in ways that distinguish them sharply from other industrial automation. A conventional industrial robot arm operates within a constrained envelope, performs a narrow set of movements, and can be designed so that its highest-wear components are easily accessible and modular. A humanoid robot moves through space on two legs, manipulates objects with multi-jointed arms and hands, and carries its power and computing systems in a chassis roughly the size and shape of a person. The number of moving parts is orders of magnitude higher. The exposure to varied mechanical stress is substantially greater.
The actuators — the motors and drive systems that move each joint — are particularly critical. Modern humanoid robots typically use a mix of electric motors with harmonic drives or planetary gearboxes, sometimes combined with hydraulic or pneumatic assist for higher-load joints. These components wear. Harmonic drives, widely used for their compact size and low backlash, are known to experience wear-related performance degradation over time, particularly under the variable loads that come with unstructured real-world movement. The exact service life of actuators in humanoid robots operating in real deployment conditions is not publicly documented by any major manufacturer — partly because the data does not yet exist at meaningful scale, and partly because companies have little incentive to publish figures that might unfavourably frame their total cost of ownership.
Beyond actuators, the mechanical demands of bipedal locomotion impose specific stress patterns that wheeled or fixed-base robots do not face. Every step distributes impact forces through ankle, knee, and hip joints. Walking on uneven surfaces, recovering from stumbles, and the low-speed postural corrections that keep a humanoid upright all generate cyclical loading that adds up over time. Bearing wear, structural fatigue, and seal degradation in the legs of a robot that takes tens of thousands of steps per shift will occur — the question is when and in what sequence.
The Technician Problem
Even setting aside how often humanoid robots will need service, there is a more immediate problem: the people qualified to perform that service barely exist yet as a workforce category.
Repairing a conventional industrial robot arm requires a trained technician, but that training is well-established. Manufacturers like FANUC, Kuka, and ABB have decades of experience running certification programmes, building parts networks, and deploying field service organisations. A factory maintenance team in a mature industrial setting typically has at least one person with the credentials and access to service common robot platforms. The ecosystem is functional because it has had time to develop.
Humanoid robots are a new category, and there is no equivalent ecosystem yet. The companies building these robots — Agility, Figure, 1X, Apptronik, and others — are primarily engineering organisations focused on development. Their field service capabilities reflect that: small specialist teams, typically based near headquarters, capable of supporting the early pilot deployments currently underway, but not yet scaled to serve a large distributed installed base. The certification programmes, spare parts networks, and independent service providers that would be needed to support humanoid robots the way industrial robot arms are supported today do not exist in any meaningful form.
This matters because the economics of maintenance depend heavily on response time. A robot that is down for two days waiting for a field service technician to fly in from another city is a different operational proposition from one where a qualified technician is available on-site or within a few hours' drive. For current deployments, manufacturers are typically providing hands-on support as part of the pilot agreement — effectively embedding service capability in the customer's facility at their own cost. This is sustainable for a handful of showcase deployments. It is not a model that scales to thousands of units across hundreds of locations.
What Industrial Precedent Suggests
The closest existing precedent for the challenge humanoid robot makers face is the early deployment of industrial robot arms in manufacturing, beginning in the 1960s and accelerating through the 1980s. That period offers some useful lessons, though the analogy has clear limits.
Early industrial robot adopters faced a version of the same problem: specialised hardware, limited service infrastructure, and a workforce with no prior experience maintaining the technology. The response, over several decades, was a combination of manufacturer-led training programmes, community college and vocational curricula adapted to industrial robotics, and the gradual emergence of independent service providers as the installed base grew large enough to sustain them. This process took a long time and was enabled by the relative simplicity of early industrial robots compared to what humanoid systems involve today.
The humanoid case is more complicated for several reasons. The mechanical complexity is higher. The variation between manufacturers is greater — a technician trained on Agility's Digit has not necessarily acquired skills that transfer cleanly to Figure's robot or to 1X's platform. And the pace of product iteration is faster, meaning that service knowledge acquired for one generation of hardware may be partially obsolete when the next generation ships. Building a durable, scalable service workforce under these conditions is a harder problem than it was for the relatively standardised platforms of the early industrial robot era.
There is also a software dimension that was not present in the same way for early industrial robots. Modern humanoid systems are deeply software-dependent: actuator control, sensor fusion, task planning, and safety monitoring all run on software that can develop faults independently of the physical hardware. A technician dispatched to address a robot that has stopped working correctly may find a hardware fault, a software bug, a firmware issue, or some interaction between all three. Diagnosing and resolving faults in systems of this complexity requires a broader and more varied skill set than maintaining purely mechanical equipment.
The Downtime Cost Question
Coverage of humanoid deployment economics tends to focus on capital cost: how much does the robot cost to buy or lease, and how does that compare to the labour it is replacing? This is a reasonable starting point, but it understates the importance of operational availability.
A robot with a purchase price of $100,000 that achieves 90% uptime over its service life is a substantially different proposition from a robot with the same purchase price that achieves 70% uptime. The difference in effective productive hours over three or five years is large. The costs of downtime — lost output, the need to maintain backup capacity or human coverage, the administrative overhead of managing service calls — add significantly to the total cost of operation in ways that capital cost figures do not capture.
Published uptime figures for humanoid robots in commercial deployment are essentially nonexistent. Manufacturers do not publish them. Customer companies operating pilots have generally not disclosed detailed operational metrics. The absence of this data is not neutral: it makes it difficult to evaluate the true economics of humanoid deployment and allows companies to present their technology in the most favourable light available. Until independent, third-party operational data starts to emerge from deployments at sufficient scale and duration, the uptime question will remain genuinely open.
Design for Serviceability
Some humanoid robot companies appear to be designing serviceability into their hardware more deliberately than others — though the details are rarely the centrepiece of public announcements. The idea is to make high-wear components modular and accessible: an ankle actuator that can be swapped in the field in under an hour rather than requiring the robot to be shipped back to a depot, or a hand assembly that disconnects cleanly from the wrist interface without requiring disassembly of the arm. Whether these design choices are being made consistently, and how well they hold up in practice, is something the industry has not yet demonstrated at sufficient scale to evaluate.
Agility Robotics has described designing Digit with operator-accessible maintenance in mind, with the ambition that a trained facility technician — not a specialist from the manufacturer — should eventually be able to handle routine service. That aspiration is reasonable, and if realised it would address part of the technician gap. But aspirations stated during development and realities encountered during scaled deployment are often different things. The honest position is that the serviceability of current humanoid platforms in real-world deployment conditions has not been tested at the scale or duration needed to make confident claims.
The broader design challenge is that humanoid robots face a genuine tension between physical performance and maintainability. The densely packed chassis that allows a robot to move with something approaching human range of motion also tends to make individual components harder to reach. Joints that are optimised for dynamic performance are not always optimised for ease of disassembly. Managing this tension — building a system that is both capable and repairable — is a real engineering constraint, not something that improves automatically as the technology matures.
What to Watch For
The maintenance and serviceability question will become harder to ignore as humanoid deployments expand beyond tightly supervised pilot programmes. A few indicators are worth tracking.
The first is whether any of the major humanoid companies begin publishing or disclosing meaningful operational data — uptime rates, mean time between failures, service call frequency — from their existing deployments. The absence of this data in current public communications is itself informative. Companies with strong operational results have reasons to share them.
The second is whether independent service providers begin entering the humanoid robot maintenance market. The emergence of third-party service companies would signal that the installed base has grown large enough to sustain them — and would provide a data point on whether the skills required for humanoid maintenance can be distributed beyond manufacturer teams.
The third is how manufacturers structure their service agreements as they move from pilot to commercial deployment. The terms of those agreements — response time commitments, uptime guarantees, parts availability, liability for downtime costs — will reveal how confident the companies themselves are in the reliability of their hardware, and where they see the remaining risks. The robot on the demonstration stage looks capable. The service contract tells you what the company actually believes about what happens next.