Created in 2025, Hexagon Robotics aims to develop humanoid robots intended for industry. Its president, Arnaud Robert, discusses the launch of his first humanoid robot, AEON, tested at BMW.
Founded last year in Sweden, Hexagon AB has established itself as a leading global player in measurement technologies and digital twins. Historically anchored in industrial metrology and positioning systems, the company has gradually broadened its scope to artificial intelligence and autonomy by creating Hexagon Robotics last year.
JDN. The AEON robot is your first product. Can you introduce it to us?
Arnaud Robert. AEON was launched in June 2025. It measures 1.65 meters and weighs 60 kilos and has 34 to 40 degrees of freedom as well as 20 onboard sensors. Its first category of use is handling: pick-and-place, sorting, moving boxes. AEON is also capable of carrying out inspections: thanks to its suite of sensors, it can inspect rooms or environments and detect defects in the infrastructure. It can also move around a factory or warehouse and generate a digital twin of the environment
Why did you choose to develop a humanoid robot rather than another type of robot?
First, customers want versatile robots, capable of operating in varied environments. Then, customers don’t want to modify their factories. They have invested hundreds of millions of euros in infrastructure designed for humans. The expectation is therefore that robots can integrate into these environments without major transformation.
What do they bring compared to human beings?
We must distinguish two logics: responding to the labor shortage and increasing the capacities of workers. The objective is in particular to delegate repetitive tasks in order to free up time for activities with higher added value. On certain tasks such as inspection or data capture, robots are already comparable, or even superior to humans, in particular thanks to machine learning capable of detecting very fine defects. On the other hand, when it comes to manipulation, humans remain superior in terms of dexterity and speed. But manufacturers are thinking more and more in terms of the production cycle. A robot capable of operating 24 hours a day can thus achieve overall productivity comparable to, or even greater than, that of a human.
What sets AEON apart from other robots of the same type?
A key point is our sensor suite. It not only allows inspection and data capture, but above all what we call spatial intelligence and spatial perception. The robot has real-time awareness of its global environment. Where other solutions focus only on the task at hand, in industrial environments it is essential to detect human movements and objects being moved, for efficiency and safety reasons.
The design of the robot is also original, notably with wheels, interchangeable hands and an automated battery changing system. What were your design choices?
We made several structuring choices. First, the wheels allow very efficient locomotion. It’s not just about going from point A to point B, but also being able to reposition itself quickly: the robot can rotate 180° or 360° in a few seconds, which is essential in industrial environments. Then, we opted for a modular approach at the effector level. Rather than designing a universal hand, we use the tool best suited to each task, which can be easily replaced as needed.
Finally, we have developed an automatic battery replacement system. Most competing robots operate on a cycle of two hours of work followed by two hours of charging, or about 12 hours of activity in a 24-hour day. Our robot can operate up to 23.5 hours per day: it simply travels to a station, replaces its battery in less than 30 seconds, then resumes its activity.
You have signed a partnership with BMW. What does it consist of?
We started by conducting tests in their factory. The two main use cases, defined by their teams, are the assembly of high-voltage batteries and production management, in particular by transferring operations from one line to another using machine learning. We are now entering the next phase: by the end of the year we want to move to full production at the BMW plant in Leipzig. The two use cases tested will be industrialized over the next six to seven months.
What lessons have you learned so far?
Data collected directly in the factory has a decisive impact. We mainly invested time in capturing the real environment and working practices. The performance gain obtained was greater than expected. Then, scaling up is a real challenge. During a pilot, we forget that the factory is a living environment. For example, if five robots work on the same line, you need to understand how they interact without colliding. The short-term objective is to increase to a few hundred units in 2027 in order to validate and prepare for scale-up.
How much do robots cost?
It is not yet communicated, but it is in line with the market, around $120,000 initially, with an expected drop as volumes increase. In terms of production, 2027 constitutes a key milestone: a few hundred units are expected by then, then several thousand per year by 2030. The main obstacle is not production, but the pace of deployment to customers. Manufacturers will not deploy thousands of robots immediately: this must be done gradually and thoughtfully.
Regarding software and robot intelligence, you mentioned partnerships with Microsoft and NVIDIA. Can you explain how you use them?
With Microsoft, it is a two-tier partnership: computing capabilities and infrastructure, as well as the imitation learning chain. The objective is to structure the entire pipeline, from data to execution by the robot, including model training, in order to optimize the whole thing from end to end. We develop our own AI models, but we rely on Microsoft for the orchestration of this chain.
With NVIDIA, we work on three levels. First for training the models, via their GPUs. Then on simulation, with virtual environments and reinforcement learning, supported by very precise digital twins. Finally, on edge AI, that is to say the AI embedded directly in the robot, which makes it possible to process sensor fusion, perception and spatial intelligence in real time.
Regarding AI models, do you train your own models?
At this point, all AI models are ours. We sometimes use open source bricks, which we heavily adjust. The only area where we are considering partnerships concerns vision-language-action (VLA) models and world models, areas of high investment. We believe it is more relevant to collaborate and then adapt these models to our needs, rather than developing them entirely in-house.
And regarding the hardware, how do you work?
We develop part of the hardware in-house, notably the “skeleton” of the robot. We also partnered with Maxon for actuators, a key component. For sensors, we work with OEM suppliers: we buy the hardware, then we develop the associated software layers ourselves. We therefore follow a hybrid approach: develop differentiating elements, purchase standard components and establish partnerships when this provides a strategic advantage.
