Humanoids can walk, sort components, manipulate objects and execute increasingly complex industrial tasks. Videos of robots dancing, running or folding laundry continue to attract millions of views.

But industrial buyers are beginning to ask a much less glamorous question: Where is the ROI?

As the sector moves from prototypes to production environments, the benchmark is changing. A successful demonstration is no longer enough. Humanoid manufacturers now need to prove uptime, throughput, safety, integration costs and, ultimately, that somebody is willing to pay for the work their machines perform.

The evidence available in September 2026 suggests that this transition has started. But only a small number of companies can currently point to meaningful commercial deployments.

Agility Robotics: perhaps the clearest commercial case

If the criterion is simply “Is a customer paying to put humanoids to work?”, Agility Robotics is one of the strongest cases today.

Its Digit robot has been operating commercially with GXO Logistics since 2024 under a multi-year Robots-as-a-Service agreement. Digit performs a deliberately unglamorous logistics workflow: moving totes from autonomous robots and placing them onto conveyors.

That is precisely what makes the deployment interesting. It is repetitive work occurring inside an existing warehouse process rather than a carefully staged robotics demonstration.

Agility subsequently reported that Digit had moved more than 100,000 totes at GXO’s Flowery Branch facility.

The company has since expanded its commercial footprint. Toyota Motor Manufacturing Canada signed a commercial agreement in February 2026 following a pilot, while Agility says its robots are operating or being introduced with industrial groups including GXO, Toyota and Schaeffler.

More importantly, Agility disclosed in 2026 that it had secured more than $300 million in multi-year contracted orders for Digit v5.

That does not mean $300 million has already been recognized as revenue. Orders, deployments and recognized revenue are three very different metrics. But it is one of the clearest signs yet that industrial customers are prepared to commit significant budgets to humanoid automation.

In financial disclosures surrounding its planned public listing, Agility presented an illustrative model in which a Digit v5 could cost approximately $200,000, plus around $20,000 for deployment and approximately $36,000 per year for software and maintenance. Using a five-year operating life, that implies a total customer cost around $400,000.

The fundamental industrial question therefore becomes measurable:

Can one humanoid create more than roughly $80,000 per year of economic value after integration, supervision and operational constraints?

If it can, humanoid robotics starts becoming an automation investment rather than a research experiment.

The winner of humanoid robotics may not be
the company with the most impressive robot.
It may be the company that can prove its robot
creates more economic value than it costs to deploy.

Figure: impressive production data, but different economics

Figure has produced perhaps the most important independently confirmed industrial dataset in the sector.

During its Figure 02 deployment at BMW’s Spartanburg plant, the robot worked on an active automotive production line handling sheet metal components. BMW says the deployment accumulated approximately 1,250 operating hours, more than 90,000 components handled, around 1.2 million robot steps and more than 30,000 BMW X3 vehicles supported.

This matters because industrial robotics ultimately lives or dies on repetition. Performing a manipulation once proves capability. Performing it tens of thousands of times begins to demonstrate reliability.

BMW has since moved to Figure 03 and a more complex logistics sequencing application, where robots must select components from containers and organize them for just-in-sequence production.

Figure therefore appears to have crossed an important technical threshold: its humanoids are no longer confined to the laboratory.

But an important distinction remains: operational success is not the same thing as demonstrated profitability.

Figure is privately held and does not disclose detailed humanoid revenue, gross margins, deployment costs or customer economics. We know considerably more about what its robots can accomplish than about how much money those deployments generate. That distinction will become increasingly important.

China is selling humanoids, but who is buying them?

China presents a completely different picture. Chinese manufacturers are rapidly industrializing humanoid production, while competition is pushing hardware prices downward.

UBTECH is particularly interesting because, unlike many private Western humanoid startups, it publishes financial results. For the first half of 2026, UBTECH reported approximately RMB 590 million in revenue from full-size embodied intelligence humanoid robot products and services, with 921 units sold.

That makes UBTECH one of the few companies where significant humanoid-related revenue can actually be observed.

But revenue needs context. A significant portion of Chinese humanoid demand has been driven by government procurement, state-backed training centres and subsidies rather than conventional private-sector automation demand.

The distinction matters. Selling a robot is not necessarily evidence that the robot is economically replacing or augmenting human labour in a factory. A machine purchased by a research laboratory or government-backed robotics training centre generates revenue for the manufacturer, but it does not demonstrate industrial ROI.

China may therefore currently be the world’s fastest-growing humanoid hardware market without yet being the world’s most mature humanoid labour market.

Tesla: enormous ambition, limited external evidence

Tesla remains impossible to ignore. Optimus potentially benefits from something many robotics startups lack: factories, manufacturing expertise, computing infrastructure and an enormous internal environment in which robots can be developed and deployed. In theory, Tesla can become both manufacturer and first customer.

But from an industrial procurement perspective, the evidence remains limited. Tesla has demonstrated increasingly sophisticated Optimus capabilities and has discussed internal factory deployment at scale. Yet detailed, independently verified figures covering uptime, task throughput, cost per operation or external customer ROI remain scarce.

The interesting question is therefore no longer whether Optimus can perform useful tasks. It is:

When will Tesla publish the equivalent of BMW’s 90,000-part Figure deployment or GXO’s 100,000-tote Digit deployment?

Until then, Optimus remains one of the industry’s most important technology programmes, but not yet its clearest commercial benchmark.

Moving 100,000 totes may be less
spectacular than a humanoid dancing.
Economically, it is far more interesting.

Apptronik, Boston Dynamics and the rest of the field

The same distinction applies elsewhere. Apptronik has been working with Mercedes-Benz on Apollo humanoid applications. Boston Dynamics and Hyundai are preparing Atlas for industrial environments. Numerous Chinese manufacturers, including Unitree, AgiBot and Fourier Intelligence, are rapidly improving hardware and reducing costs.

These programmes matter. But pilots, partnerships and purchase announcements should not be confused with scaled commercial operations.

For technical buyers, five metrics are becoming much more useful than demo videos:

• Autonomous operating hours
• Intervention rate
• Task throughput
• Total deployment cost
• Cost per completed operation

Add safety certification, maintenance requirements and integration with factory and warehouse management systems, and the real difficulty of humanoid deployment becomes apparent.

The hidden metric: human intervention

Perhaps the most important number in humanoid robotics is one companies rarely advertise: how often does a human need to help the robot?

A humanoid could theoretically work 20 hours per day. But if remote operators frequently intervene, reset failed tasks or manually control difficult manipulations, the economics change dramatically.

This is why autonomy percentage alone can be misleading. A commercially meaningful humanoid needs not only high autonomy but also a sufficiently low intervention frequency and short recovery time.

A robot completing 99% of actions autonomously sounds impressive. But in a workflow containing thousands of actions per shift, the remaining 1% can still generate substantial human supervision.

The industry’s next transparency battle may therefore concern intervention rates rather than walking speed or dexterity.

Why warehouses and factories are winning first

The earliest viable humanoid applications share several characteristics. They occur in relatively structured environments. The workflows repeat frequently. Objects and infrastructure are reasonably predictable. The economic value of automation can be measured.

Labour shortages, ergonomics and undesirable repetitive work also create an existing business problem. This explains why logistics and automotive manufacturing are emerging ahead of household robotics.

The humanoid’s theoretical advantage is not necessarily that it can outperform a specialized industrial robot. Often it cannot. Its advantage is that it may eventually operate inside infrastructure designed for humans without requiring the facility to be rebuilt around automation.

Doors, shelves, carts, workstations and production lines were designed around human geometry. A sufficiently capable general-purpose robot could exploit that installed infrastructure. That is the economic hypothesis behind the humanoid.

The real competitor isn’t another humanoid

There is another uncomfortable question the industry needs to answer. A humanoid robot is not competing only against Figure, Digit, Optimus or Apollo. It competes against an industrial robot arm, an autonomous mobile robot, a conveyor, machine vision, a purpose-built automation cell and, of course, a human worker.

If a $50,000 automation solution can perform the same workflow more reliably than a $200,000 humanoid, there is little economic reason to use a humanoid.

The humanoid only wins when its generality creates value. That could come from performing several tasks with the same hardware, being redeployed without major infrastructure changes, handling environments designed for humans, or learning new workflows through software rather than mechanical reengineering.

This is why the ultimate commercial metric may not be cost per robot. It could be cost per task across the robot’s lifetime.

From cost per robot to cost per productive hour

Industrial buyers may soon stop asking: How much does the humanoid cost?

The better question is: How much does one productive autonomous hour cost?

Imagine a $200,000 robot operating reliably for five years. If it works only a few hours per day, requires frequent supervision and performs one narrow workflow, the economics may be poor. If the same hardware works across multiple shifts, performs several tasks and learns additional workflows through software updates, its economics change dramatically.

The real value is therefore not simply making robots more capable. It is allowing the same physical asset to become more productive over time through software. That is fundamentally different from traditional fixed automation.

So, who is actually making money?

As of September 2026, the answer requires nuance.

Agility Robotics has one of the strongest cases for genuine paid industrial deployment and has disclosed a substantial contracted order book.

Figure has some of the strongest publicly verified production evidence, particularly through BMW, but its underlying revenue and deployment economics remain private.

UBTECH is generating substantial identifiable humanoid-related revenue, although the Chinese market’s dependence on public procurement and training infrastructure complicates comparisons with Western industrial deployments.

Tesla, Apptronik, Boston Dynamics and many emerging competitors remain strategically important but have yet to publish enough standardized commercial data to make direct ROI comparisons possible.

Nobody has yet demonstrated, publicly and at massive scale, that general-purpose humanoids can economically replace conventional automation or human labour across multiple industries.

But something important has changed. Two years ago, the central question was: Can humanoid robots work? In 2026, there is enough evidence to answer yes, under specific conditions.

The next question is much harder: Can they work cheaply enough, reliably enough and autonomously enough to generate sustainable profit?

That is the race that matters now.

The winner of humanoid robotics may not be the company with the robot that runs fastest, looks most human or produces the most impressive demonstration. It may simply be the company that can prove one equation:

Value of work performed > total cost of deploying the robot.

Everything else is still a demo.

Cet article Humanoid Robots: Enough Demos. Who Is Actually Making Money? est apparu en premier sur Robot Magazine.

For the robotics sector, this evolution is particularly significant. Artificial intelligence is no longer limited to generative models capable of producing text, code, images or video. It is increasingly moving beyond screens and into systems that interact with the physical world.

More autonomous industrial robots, machines capable of understanding their surroundings, intelligent mobile systems, autonomous vehicles and humanoid robots are all part of a broader convergence between AI and physical systems.

That convergence is one of the areas AI Everything Abu Dhabi intends to highlight this year.

A global event focused on applied AI

According to the organisers, AI Everything Abu Dhabi is expected to bring together more than 350 AI companies, startups and international technology players from over 60 countries.

The event is also set to host more than 200 investors managing approximately US$50 billion in assets, alongside more than 150 AI-focused startups.

The event therefore goes beyond the format of a conventional artificial intelligence conference.

Its ambition is to showcase technologies already being deployed in fields such as government, finance, healthcare, energy and manufacturing, with a strong emphasis on practical applications and real-world use cases.

This focus on applied AI reflects a broader shift in the market.

After several years dominated by the race to build increasingly powerful large language models, companies are now looking at how AI can be turned into systems that directly improve operations, industrial processes and infrastructure.

From generative AI to Physical AI

Among the six main technology areas highlighted by AI Everything Abu Dhabi is Physical AI, alongside robotics and autonomous mobility.

Other focus areas include intelligent infrastructure and new energy, frontier AI models, sovereign AI, enterprise and agentic AI, as well as sovereign security and quantum technologies.

The presence of Physical AI in this programme is particularly noteworthy.

Physical AI generally refers to artificial intelligence systems capable of perceiving, understanding and acting within a physical environment.

For a robot, the challenge is therefore no longer simply to generate the correct answer to an instruction.

A machine must be able to interpret its surroundings using cameras, sensors or other perception systems, understand a task, make a decision and then carry out an action in the real world.

This combination of perception, reasoning and action is now becoming one of the key development areas in advanced robotics.

It is particularly relevant to humanoid robots, autonomous mobile robots, intelligent industrial machines and robotic manipulation systems.

Robotics is entering the era of AI models

One of the major transformations currently taking place in robotics is the emergence of artificial intelligence models capable of processing several types of information at the same time.

Vision, language, audio, sensor data and human instructions can increasingly be brought together within a single architecture.

This could lead to robots that are less dependent on highly specific programming for every individual task.

Instead of following only predefined sequences, some machines can be trained to interpret instructions and adapt their behaviour according to their environment.

For industry, this represents a significant change.

Traditional robotics remains extremely effective in structured environments such as automotive welding, assembly, palletising, painting, material handling and quality control.

New generations of autonomous systems, however, are increasingly being designed to operate in much more variable and unpredictable environments.

This is where Physical AI becomes particularly important.

Agentic AI could also reshape robotic systems

AI Everything Abu Dhabi will also focus on another rapidly developing field: Agentic AI.

The event is set to feature workshops covering agentic enterprise systems, multi-agent orchestration, responsible AI, Physical AI and rapid prototyping.

AI agents are designed to go beyond simply generating a response to a prompt.

They can receive an objective, plan multiple steps, use digital tools and coordinate different actions.

In robotics, this approach opens up interesting possibilities.

An AI agent could, for example, organise a logistics operation, distribute instructions to multiple autonomous mobile robots, monitor progress and modify the plan if conditions change.

Over time, the boundary between software agents and autonomous machines may become increasingly difficult to define.

Infrastructure is becoming strategic for intelligent robotics

This new generation of systems does not depend on robots alone.

Computing power, data centres, cloud platforms, semiconductors, networks and energy infrastructure are becoming essential components of the ecosystem.

AI Everything Abu Dhabi is designed to connect these different technology layers, bringing together data infrastructure, compute capacity, energy, agentic AI, enterprise automation, quantum technologies and sovereign security.

This convergence matters greatly for robotics.

An autonomous robot may need to process certain tasks locally in order to minimise latency, while other operations, such as model training or large-scale data analysis, may require far more powerful computing infrastructure.

The next generation of intelligent machines will therefore depend not only on progress in robotics, but also on advances in processors, AI models and computing infrastructure.

Microsoft, Intel, AWS, Mistral AI and other major technology players expected

The event is set to welcome a number of major international technology companies.

Exhibitors announced include AWS, Alibaba Cloud, ASUS, Cloudflare, Dell, HPE, Intel, Microsoft, Mistral AI, Nokia, Nebius, Nutanix, Red Hat, Snowflake, SUSE and Zoho.

Although not all of these companies are directly involved in robot manufacturing, their technologies form part of the infrastructure on which modern AI increasingly depends.

The development of more AI-driven robotics relies heavily on computing capacity, cloud services, edge computing, cybersecurity and data management.

AI Everything Abu Dhabi therefore aims to bring together the different layers of this emerging technology stack within the same event.

Specialised AI models are becoming increasingly important

The event also aims to highlight the growing diversification of the artificial intelligence model landscape.

The organisers mention companies developing Arabic large language models, multimodal AI, Voice AI technologies, open-weight models and enterprise agents.

This diversification could have important implications for robotics.

Not every robot will require an enormous model containing hundreds of billions of parameters.

In many cases, industrial users may prefer smaller, specialised models capable of operating directly on a machine or within local infrastructure.

The development of more efficient models could therefore become a key factor in the wider adoption of Physical AI.

Among the companies highlighted by the event is ModelBest, creator of the MiniCPM family of relatively compact AI models.

Abu Dhabi accelerates its AI strategy

AI Everything Abu Dhabi also has a strong institutional dimension.

The Department of Government Enablement – Abu Dhabi (DGE) is presented as the Host Partner, while a number of UAE technology and academic institutions are expected to participate.

These include the Advanced Technology Research Council, the Technology Innovation Institute, the UAE Cybersecurity Council and Khalifa University.

One of the ambitions highlighted by the organisers is to demonstrate how Abu Dhabi intends to build government services and public administration increasingly around artificial intelligence.

Beyond the public sector, the UAE is also seeking to develop a wider ecosystem bringing together research, investment, computing infrastructure and technology companies.

For robotics players, this dynamic could create new opportunities in autonomous systems, industry, logistics and intelligent infrastructure.

An event worth watching for the robotics ecosystem

For Robot Magazine, AI Everything Abu Dhabi is therefore relevant far beyond the field of generative AI.

The inclusion of Physical AI, robotics, autonomous mobility, agentic AI and advanced computing infrastructure illustrates how quickly the boundaries between artificial intelligence and physical machines are beginning to disappear.

For many years, software and robotics developed as relatively separate technology domains.

That distinction is becoming much less clear.

Artificial intelligence models are increasingly involved directly in the perception, planning and decision-making processes of autonomous systems.

The key question is therefore no longer simply how far AI models will progress.

It is also what those models will be able to do once they are integrated into machines capable of acting in the physical world.

AI Everything Abu Dhabi 2026 could provide an important glimpse into that next phase.

Cet article AI Everything Abu Dhabi 2026: Physical AI, Robotics and Autonomous Systems Take Center Stage est apparu en premier sur Robot Magazine.

The arrival of LLMs (Large Language Models) such as GPT, Claude, Llama, or Gemini opens a new era: that of agentic robotics. These models allow AI agents to interpret instructions in natural language, plan complex actions, and orchestrate multiple systems in real time. In other words, robots gain a brain that is far more flexible and proactive.

This article explores how LLMs are transforming robotics, real-world examples already underway, and what this means for the future of industry and society.

What is agentic AI applied to robots?

Agentic AI refers to an approach where artificial intelligence does not just respond to one-off queries but acts autonomously to achieve a goal. An agent can:

• perceive its environment through sensors or external data,
• reason using a model like an LLM,
• plan steps to reach a goal,
• act by controlling a robot or interacting with other systems.

In the robotic context, this means machines no longer just repeat preprogrammed tasks: they become capable of adapting to new situations, learning, and even collaborating with one another.

LLMs as the engine of autonomy

Why do LLMs change the game?

• Natural language understanding: an operator can give a simple order—“prepare this order and check the quality of the parts”—and the agent translates it into concrete robotic instructions.
• Sequential reasoning: LLMs can break down a complex task into logical steps. For example, a logistics robot must first locate the item, avoid obstacles, then transport it to the station.
• Adaptability: unlike rigid scripts, LLMs can handle the unexpected. If a path is blocked, they generate a new plan.
• Multi-system integration: they serve as an orchestration layer between multiple robots, software, and databases.

In short, the LLM becomes a kind of cognitive conductor, giving robots greater intelligence and flexibility.

Real-world examples: when LLM agents control robots

• Logistics and warehouses: Amazon Robotics and Boston Dynamics are already experimenting with AI agents that manage fleets of mobile robots using natural language commands.
• Agriculture: a robot like Ted from Naïo Technologies could, with an LLM agent, analyze weather data and adjust its weeding strategy.
• Healthcare: surgical robots could be supervised by an LLM agent capable of flagging anomalies or instantly consulting a global medical database.
• Industrial maintenance: a cobot with an LLM can receive an oral instruction—“check vibration levels on machine 4 and report if maintenance is needed”—and carry out the analysis with its sensors.
• Humanoid robots: projects like Figure AI or Tesla Optimus aim to create humanoids that can understand and execute general instructions thanks to LLM-based agents.

Technical and ethical challenges

Despite its promise, integrating LLMs into robotics raises several challenges:

• Reliability: LLMs can “hallucinate” or produce incorrect instructions. In industrial or medical contexts, errors are unacceptable.
• Safety: how can we ensure an agent doesn’t make a dangerous decision? Safeguards and human oversight remain essential.
• Energy costs: LLMs are computationally intensive, limiting their direct use onboard robots. Hybrid solutions (cloud + edge computing) are emerging.
• Legal responsibility: if a robot controlled by an AI agent causes harm, who is liable—the manufacturer, the operator, or the model provider?
• Ethics: the line between autonomy and human dependence must be clarified to prevent abuse or loss of control.

The future: towards “collectives of robotic agents”

The convergence of LLMs, connected sensors, and robotics paves the way for a future where robots operate in intelligent ecosystems. Imagine:

• A factory where every robot is an autonomous agent communicating with others via a natural language protocol.
• A fully automated farm where agricultural robots self-organize to optimize harvests according to weather and soil.
• Hospitals where assistance, transport, and surgical robots coordinate under the supervision of specialized AI agents.

These “agentic collectives” could transform productivity, reduce costs, and above all, open new forms of human-machine collaboration.

The union of agentic AI and robotics powered by LLMs marks a major turning point. Robots are no longer simple preprogrammed executors: they become partners capable of understanding, reasoning, and acting proactively.

For industry, it is an opportunity to gain flexibility and efficiency. For society, it is both a chance and a challenge to redefine the role of humans in a world where machines and intelligent agents coexist.

As with most technological revolutions, the key will lie in balancing autonomy and control, innovation and regulation. But one thing is certain: the era of agentic robots powered by LLMs is only just beginning.

FAQ: Agentic AI and Robotics

Cet article Agentic AI and Robotics : When LLMs Give Robots a Brain est apparu en premier sur Robot Magazine.