In early July, more than 50 robots were deployed in Davos for the Tech Summit. For Marco Pavone, senior director of AI Research at Nvidia and a professor at Stanford, quoted by the magazine PME, “we are entering the era of physical AI, where intelligence is no longer confined to the digital world, but understands, reasons and acts in the physical world”. In its special report on the subject, PME asks a simple question: can Switzerland matter in this race, against a China that is already mass-producing and a United States betting on its own AI models?
The context calls for caution as much as attention. Projections are circulating, such as Morgan Stanley’s, which mentions 1 billion humanoid robots in service worldwide in 2050. But the magazine also points out that, in practice, robots often remain slow, sometimes clumsy and far from fully autonomous.
How a robot powered by artificial intelligence works
Physical AI refers to AI capable of perceiving its environment and acting within it. Unlike conventional software, it requires several disciplines to converge: AI models, sensors, mechanics, electronics and an understanding of the environment.
Two Swiss examples described by PME show concretely what this involves. The quadruped robot ANYmal, from ANYbotics, learns to move through deep reinforcement learning: it first trains in simulation, then reproduces those movements in the real world. Used in energy, chemicals and offshore, it detects, for example, bearing defects using ultrasound or small temperature differences using thermography. The fleet collects more than 150,000 inspection points per month, which the AI sorts to spot anomalies. For now, humans keep the final say on how they are interpreted.
Flexion, founded in Zurich at the end of 2024 by, among others, former Nvidia researchers, works on the other end of the chain: the “brain”. Its Reflect platform is designed to enable a humanoid robot to understand an instruction in natural language, then plan and carry out a sequence of actions, by combining a vision-language model and reinforcement learning. The company stresses the ability to recover after a mistake, since a series of small slip-ups can cause an entire mission to fail. It raised 50 million dollars in 2025.
The strengths of Swiss robotics
For Jan Kerschgens, an innovation and deeptech expert and co-director of the Swiss Center for Digital Sovereignty, Switzerland has “an ecosystem that few countries can match”: first-rate academic research, with EPFL and ETH Zurich, combined with a precision industrial base inherited from microtechnology, watchmaking and medtech. In his view, simultaneous mastery of software and high-precision mechanics is rare, and that is precisely what physical AI demands. He places the Swiss advantage in the integration of embedded AI into industrial, medical or assistive robotics.
ANYbotics’ track record serves as a benchmark: founded in 2016, the company has more than 200 employees and has raised around 120 million francs. Other players occupy neighbouring niches, such as Ascento, an ETH Zurich spin-off specialising in security robots. On a different scale, Nicolas Fontaine, director of the Valais company Workshop 4.0, known for its Roboclette, sees safety as one of the main challenges ahead.
What China already dominates
On the Chinese side, the figures cited by PME are clear. According to McKinsey, Chinese manufacturers accounted for 57% of industrial robot installations in China in 2024, compared with around 28% ten years earlier. “China is no longer primarily a market for foreign robotics technologies; it has become a leading producer in its own right,” sums up Christian Jansen, Partner at McKinsey & Company.
The mechanism is industrial. For its robots, China reuses the motors, batteries, sensors and electronic components coming from two already massive sectors, electric vehicles and consumer electronics. In several hubs, the parts needed for a prototype can be found close by, which shortens development cycles: according to Christian Jansen, a designer can obtain a modified part in a few days rather than a few months, and run through more versions in a year. In humanoids, an industrial policy supports the multiplication of models and trials.
Christian Jansen nevertheless considers it misleading to speak of a race with a single leader: China is betting on volumes and costs, the United States on vertical integration and proprietary AI, while Europe retains a head start in certain high-precision components.
The weaknesses: moving from the lab to the market
Scientific excellence does not guarantee industrial success. Jan Kerschgens believes the building blocks exist and that the challenge is above all to “make the whole chain work”, from research to market. The weak point he identifies is growth financing: “It is neither ideas nor talent that are lacking, it is domestic growth capital.” He suggests making it easier for pension funds, which still invest little in venture capital, to access this asset class, without forcing risky choices on them. The sources do not put a figure on the funding gap between Switzerland and its competitors.
He also warns against the risk of becoming a mere user of technologies designed elsewhere. In his view, the value of tomorrow’s robots will lie as much in their AI models as in their motors and sensors, hence the importance of pursuing efforts in open AI models and digital infrastructure.
Where the game will be played
PME magazine nevertheless notes that China still lags behind in movement precision, sensors, safety electronics and above all cutting-edge AI. For Swiss companies, Christian Jansen therefore recommends not trying to compete on volumes, but betting on precision and engineering, building safety into products and selling complete subsystems rather than mere components. “Robotics is being revalued as a physical AI business rather than as a machinery business,” he observes.
The sources thus describe a Switzerland that is well placed in research and precision, but faces a challenge of industrialisation and capital; its share of the global market is not specified. It remains to be seen, as PME notes, whether China will manage to turn its hardware lead into superiority in machine intelligence: it is on this ground, according to the magazine, that Switzerland and Europe still have a card to play.
