Summary

• Agile Robots chief executive Zhaopeng Chen says the company plans to produce between 1,000 and 2,000 Agile ONE S and Agile ONE SN humanoids during 2027, with production beginning by the end of 2026.
• Agile ONE SN combines a compact humanoid body, tactile hands, Nvidia Jetson Thor computing and Isaac GR00T development tools, with availability planned for early 2027.
• Universities and corporate research laboratories are the initial market. Production targets, institutional interest and promised development efficiencies are not equivalent to confirmed orders or industrial deployment.

Agile Robots plans to manufacture between 1,000 and 2,000 compact humanoid robots during 2027, targeting universities and corporate research laboratories developing the technology needed for more adaptable physical automation.

Chief executive Zhaopeng Chen disclosed the production ambition on 21 October as the Munich-based company introduced Agile ONE SN, a humanoid reference platform developed with Nvidia and intended to provide researchers with an integrated hardware and software environment.

Production of the Agile ONE S and Agile ONE SN models is expected to begin before the end of 2026. The Nvidia-equipped reference configuration is scheduled to become available in early 2027.

The figures are manufacturing intentions rather than completed production or confirmed customer orders. Agile Robots has not disclosed a binding order book covering the proposed quantities, the selling price of the reference platform or the investment required to support the planned output.

A research platform reduces integration work

Humanoid robotics researchers need machines that allow them to test how AI models control movement, interpret their surroundings and handle physical objects.

Developing that environment can involve extensive engineering before an experiment begins. Teams may need to select mechanical hardware, connect sensing systems, configure onboard computing and integrate separate tools for simulation, training and deployment.

Those activities require expertise that may be unrelated to the particular robot learning problem a research group wants to investigate.

Agile ONE SN is intended to reduce that initial work by supplying a common machine with Nvidia’s development software already configured.

The reference platform uses Agile’s compact humanoid body, with articulated limbs and hands designed for manipulation experiments.

According to Nvidia’s product description, the robot stands approximately 1.4 metres tall and has 31 degrees of freedom in the body, with two five-finger hands providing additional movement and tactile sensing.

Each arm has a stated payload capacity of three kilograms, making the platform suitable for investigating relatively small-object manipulation rather than industrial lifting applications requiring much greater strength.

The hands incorporate force control and tactile sensing, allowing developers to examine how the robot responds when it makes contact with an object.

These capabilities can support experiments involving grasping, positioning and handling materials whose shape or movement cannot be represented reliably through visual information alone.

However, the usefulness of the platform depends on how consistently those functions perform and whether learned behaviours remain reliable outside the conditions used during training.

Onboard computing supports local control

Agile ONE SN incorporates Nvidia Jetson Thor computing, allowing models and control software to operate directly on the robot.

Local processing can reduce dependence on sending every sensor observation to a distant computing facility before the robot decides how to move.

The system also includes cameras, lidar and motion sensors that provide information about nearby objects, the robot’s own movement and the environment in which it operates.

Combining those inputs allows developers to investigate how a machine should respond when conditions change, rather than relying exclusively on fixed instructions prepared for one precise arrangement of objects.

The Nvidia Isaac GR00T platform provides AI models and development tools supporting the progression from collecting examples to training and evaluating robot behaviour.

Researchers can use simulation to explore potential movements before transferring selected policies to the physical platform.

However, a movement that succeeds in simulation may fail on real equipment because of differences in friction, balance, sensor behaviour or the characteristics of objects being handled.

Physical validation therefore remains necessary, particularly where an incorrect movement could damage equipment or endanger a person nearby.

The companies are planning integration with Nvidia Halos safety technology, but that work has not been presented as completed certification of Agile ONE SN for unsupervised industrial operation.

Universities form the initial customer market

Agile Robots is positioning the new machines principally for academic and corporate research.

Stanford University and ETH Zurich are among the institutions identified as potential users, although neither the company nor Nvidia has disclosed confirmed purchase agreements with them.

Research laboratories may value a common hardware configuration because it allows experiments to be reproduced or compared between teams using similar machines.

Where laboratories develop different control software for the same platform, their results may be easier to evaluate than experiments conducted on robots with substantially different physical characteristics.

Common configurations can also support shared development resources and reduce the need for researchers to create separate software interfaces for every machine.

Nvidia says the integrated design can reduce the initial setup process from weeks to days. In the accompanying interview, the company’s robotics representatives described potentially avoiding months of integration work associated with purchasing separate components.

These are estimates of preparation effort, not independently verified reductions in the time required to develop a reliable industrial robot skill.

A laboratory could begin collecting training data more quickly while still needing substantial experimentation before its robot can complete a complex task consistently.

The difference between development convenience and actual operational performance will be important when potential customers assess the system’s commercial value.

Production plans test manufacturing capability

Manufacturing 1,000 to 2,000 machines during 2027 would require Agile Robots to move beyond building individual research platforms and demonstrate more repeatable assembly and quality control.

Humanoid robots contain motors, gear mechanisms, sensors, computing systems and structural components that must operate together within defined tolerances.

Maintaining consistency across a substantial production run requires reliable components, calibration procedures and testing capable of identifying defects before machines are delivered.

Even small variations in mechanical behaviour can affect the performance of software developed using a reference platform, reducing the advantage of providing researchers with a common configuration.

Customers will also need maintenance arrangements, documentation and access to replacement components as machines are used for experiments that may place unusual demands on their hardware.

Production quantities alone will not establish financial success. The company must generate sufficient demand and revenue to support manufacturing, technical assistance and future product development.

Agile Robots has not published confirmed manufacturing capacity, customer deposits or prices that would allow its 2027 production target to be compared with contracted demand.

Industrial applications remain a longer-term objective

Agile Robots already develops industrial automation equipment, including larger robotic systems intended for manufacturing and logistics.

The compact research platforms could support the development of perception, manipulation and movement capabilities that eventually transfer to machines working in commercial environments.

That transfer is not automatic. Differences in weight, strength, joint arrangements and operating conditions can change how a learned behaviour performs.

A robot trained to handle lightweight objects on a compact platform may require substantial additional development before the same behaviour can be used safely with heavier equipment.

Europe’s wider robotics sector includes companies pursuing different approaches to physical AI. Techopia previously examined NEURA Robotics’ multimodal model, which combines several forms of perception and memory to support physical tasks.

Agile Robots’ current development concerns the availability and planned manufacturing of a research platform rather than a new claim about the performance of NEURA’s model or another competing system.

The commercial comparison for humanoid robotics also depends on conventional automation. Fixed-purpose industrial robots may remain more economical for repetitive work, while adaptable machines could become attractive where tasks vary too frequently for extensive dedicated programming.

Factories considering future humanoid deployments will need evidence of task completion rates, operating costs, maintenance, recovery from faults and safety under working conditions.

Agile Robots’ immediate objective is to supply researchers with machines that help generate such evidence and develop reusable control methods.

The first scheduled production and delivery milestones will test whether the company can manufacture a consistent platform at the intended volume. Wider industrial adoption will depend on whether the capabilities developed using those robots subsequently demonstrate reliable performance and acceptable economics in commercial environments.