Are Humanoid Robots about to take over carmaking?

There is a particular quality to the excitement currently surrounding humanoid robots in automotive manufacturing. It has the character of a technology arriving on schedule with its own press strategy: structured announcements, carefully curated demonstration footage and language calibrated to convey both near-term commercial credibility and long-term transformational potential simultaneously. The result is a conversation that is, in equal measure, genuinely significant and genuinely difficult to read clearly.

BMW has completed what it describes as the first deployment of humanoid robots inside an active car plant, with Figure's 02 model supporting production of more than 30,000 BMW X3s at Spartanburg, South Carolina over ten months, operating in ten-hour shifts and handling 90,000 components. BMW has since confirmed deployment at its Leipzig plant - the first instance of physical AI of this kind entering a European automotive production environment at scale.

Toyota Motor Manufacturing Canada, meanwhile, formalised the first commercialised humanoid robot deployment in Canadian automotive production following a year-long evaluation programme. Tesla's Optimus is in internal trials with commercial rollout targeted for this year. Nvidia made physical AI a centrepiece of its GTC conference in early 2026.

The volume of activity is real. So is the noise. What is less certain is whether the two are proportionate to one another.

What has actually been deployed

The deployments that have taken place are not trivial. They are structured, multi-year programmes with specific performance criteria - not demonstrations staged for the benefit of press attendance. BMW's engagement with Figure 02 at Spartanburg was a genuine production programme. The year-long evaluation Toyota conducted before a single robot went live on the Canadian line reflects methodological rigour considerably more demanding than the typical pilot.

The AMS team reported directly from BMW's competence centre for physical AI in Leipzig in late February, where the company demonstrated Aeon, its second humanoid platform developed in collaboration with Hexagon. The intent, as BMW communicated it, is for humanoids to work hand in hand with human employees on the line. What is easy to miss in the headlines, however, is that the broader challenge of deploying intelligent automation at scale - humanoid or otherwise - remains stubbornly difficult.

Andreas Kühne, Program Manager for Artificial Intelligence in Production and Logistics at Audi, oversees more than 100 AI initiatives across the company's 360 Factory strategy. His experience offers a useful corrective to the idea that deployment is simply a matter of technology readiness.

"It's easy to have a good AI idea," he observes, "and most of the time it's also easy to build a prototype that shows something in a fancy way. But the key challenge is data. You have to have the data in a certain quality, integrated and available, with a certain standard and semantic that works across every production line and every production plant. Otherwise you have to build a translator for the AI solution to each individual system - and that takes quite a lot of time."

That structural friction between good ideas and scalable reality is not unique to AI in its conventional forms. It is the condition that any new automation technology, including humanoid robotics, will have to navigate.

The two technical barriers that remain unsolved

Mike Wilson, Chief Automation Officer at the Manufacturing Technology Centre, brings more than 40 years of experience to his assessment of where humanoid robots actually stand. He is a Visiting Professor of Robotics and Automation at Loughborough University, Chair of the UK Automation Forum and General Assembly member of the International Federation of Robotics. He is also, on this particular subject, usefully sceptical.

"I am somewhat sceptical," Wilson acknowledges directly, "and I think there are still significant technical challenges that need to be addressed." He identifies two with particular precision: dexterity and safety.

On dexterity, the problem is unsolved in a way that materially limits the range of tasks humanoids can perform. "We need grippers and tools that the humanoids can use that are equivalent, in practical terms, to what we can do with our hands. I know that progress is being made very quickly, but we are not there yet. And until that one's solved, they're not going to be that general-purpose automation device."

The safety challenge is, if anything, more immediate. The architecture of industrial safety built up over decades assumes a specific and reliable response to an emergency stop signal: power is cut and the machine halts. That logic does not transfer to a bipedal humanoid. 

"If you activate the emergency stop on a humanoid, it falls over," Wilson observes. "Given that the humanoids currently being used in factories are about the same size and weight as a person, I don't particularly want one of those things falling on me."

This is not a minor engineering detail. It is a fundamental constraint on the degree to which humanoids can operate in close proximity to human workers. The consequence, visible in current deployments, is that humanoids tend to be used in areas where human workers are not simultaneously present - a workable arrangement in certain contexts, but a direct constraint on the human-robot collaboration model that BMW and others have articulated as the long-term vision.

The expertise bottleneck applies here too

Beyond the physical engineering challenges, there is a deployment problem that Wilson, Kühne and A.J. Camber, VP and Head of the Software Business Group at Solidigm, all identify from different vantage points: the gap between a technology that works in principle and one that organisations can actually operate.

Camber, whose platform Lucetta launched in early 2026 to let engineers build and iterate visual inspection models without data science expertise, frames the problem in terms that extend well beyond computer vision. "The scarcity of data science expertise is one of the main issues we see. The ease of use for an industrial or mechanical engineer on the line is just not there.

"And if you're lucky enough to have one of those data scientists, you still have this side problem where you have a dependency on just a few individuals, which can be challenging. It would be easier to build trust in the technology across the organisation if we can get more people involved."

That democratisation argument matters for humanoid robotics too. The more that deployment and operation requires specialist expertise concentrated in a small number of individuals, the slower and more fragile adoption will be.

For a technology whose long-term value proposition depends on widespread deployment across heterogeneous production environments, that is a structural constraint as significant as the dexterity problem.

Where humanoids do have a credible near-term role

None of this renders humanoid robots without value in automotive manufacturing. Wilson is precise about where viable application exists today, and the logic of that precision is worth following carefully.

The most credible near-term deployment case is the retrofitting of existing facilities. "The use cases where humanoids do have viable application today are operations that are currently manual and have not been automated using other robot types," Wilson explains. "We need a machine that is similar in size, shape and capability to a person. Unloading fixtures and moving parts to other locations can be handled successfully because you can potentially deploy the humanoid without significantly modifying the existing facility. That makes it a relatively cost-effective form of automation in a retrofit context."

This positions the humanoid not as the default form factor for the factory of the future, but as a practical tool for extending the reach of automation into spaces where conventional industrial robots cannot economically go. For manufacturers facing the combined pressure of high labour costs and significant capital requirements for full automation - particularly in Western European plants, where both are acute - this is a genuine and near-term business case.

Kühne's framing at Audi points toward the same discipline: before any use case is taken to production, it must demonstrate viability across all plants, not just one. "One of the prerequisites we insist on is a commitment from, say, the paint shop or the press shop across all production plants. Otherwise, in most cases, the use cases don't scale." Applied to humanoid deployment, that standard sets a high bar.

Where traditional automation still wins

The corollary of the retrofitting argument is equally instructive. Where a new facility is being designed from the ground up, the case for humanoids weakens considerably. "If you are building a new facility," Wilson notes, "you would design it in a way that allows automation using simpler devices such as industrial robots. That would be a far more cost-effective approach."

The area of the plant where this is most pronounced is trim and final assembly - the phase that deals with the most complex, least standardised physical tasks. "Because of issues with dexterity, it will be some time before we see humanoids working in trim and final, where they are dealing with uncontrolled parts. Trying to fit trim into a car remains a very difficult thing to automate. Humanoids do not yet have that capability."

The gap between public narrative and operational reality is most visible precisely here, in the tasks that are hardest and most valuable to automate.

Not a breakthrough - but a basis, building for the future

The appropriate frame for humanoid robots in automotive manufacturing is trajectory, not breakthrough. The technology is real, it is developing at pace and it has identified a specific and credible near-term role. What it has not yet done, and what the evidence suggests it will not do for some years, is fulfil the broader promise of a general-purpose automation device capable of performing the full range of tasks currently carried out by human workers on a modern assembly line.

Wilson puts it plainly: "In certain use cases, humanoids do have a viable operation. But the whole point of a humanoid is that it is a general-purpose automation device. We are not there yet. It is going to be a little while before we get there."

That is not a dismissal. It is a timeline. And in automotive manufacturing, where investment cycles are long and the cost of misallocated capital is real, timelines matter.