Meta reportedly built a humanoid robot with Fourier
A media report describes a first prototype. Laboratories in Zurich and Lausanne are also studying how machines grasp safely and find their way in everyday life.
Meta is reported to have built a humanoid robot together with the Chinese company Fourier and to be working on a second version already. Reuters reported this on 31 October, citing "The Information" and its informed sources. The agency was initially unable to verify the details independently. A product launch cannot yet be derived from this report.
The report lands in a field of research in which Swiss universities are also active. Machines should not only recognise objects but touch them, move them and react to changes. Perception, learning methods and mechanics have to work together. A human-like body alone does not solve these tasks.
Meta lets machines learn from videos
Meta has already documented part of its work publicly. In June 2025 the company presented V-JEPA 2, a so-called world model. It is meant to learn from observed sequences what consequences a movement is likely to have. Instead of only producing a description, such a model can thereby help plan a robot action.
According to the research report, the base model was pre-trained on more than one million hours of internet videos. For the variant coupled to actions, fewer than 62 hours of robot videos were added. The researchers deployed it on Franka robot arms in two laboratories. There, objects could be picked up and placed at a target specified by images.
These are trials with specific tasks and experimental setups. They do not prove that a robot can perform arbitrary housework. Which software the Fourier prototype now reported uses does not emerge from the short Reuters report. The published V-JEPA results therefore must not simply be read as a performance record of this device.
In Zurich, the work starts with the hand
At the Soft Robotics Lab at ETH Zurich, the team led by robotics professor Robert Katzschmann develops hands that learn movements from examples. In a setup described by ETH, people wearing a sensor glove demonstrate how they grasp a bottle. Cameras record the movement. From such data the system learns to handle objects it has not seen before.
The hand has 21 degrees of freedom, 28 together with the arm. This refers to the movement possibilities that can be controlled separately. The ETH spin-off Mimic Robotics emerged from the research in 2024 and pursues applications in manufacturing and logistics. Swiss value creation can thus also arise in a specialised hand or its control, not only in building a complete humanoid.
A more recent ETH project shows how closely locomotion and grasping are linked. In "Fingers as Legs", a robotic hand carries its own power supply and computer and walks on its fingers. When it lifts one finger for a step or a key press, the others have to carry its weight. The researchers first train the movements in a simulation.
The project page also names limits: separate control models are used for different abilities. Before the keyboard trials, the hand is aligned manually. A demonstrated key press is therefore not yet an independent understanding of a computer task. In the get-up trials shown, the movement succeeded in 21 of 25 cases. That is a result of this test, not a success rate for other grips or a complete household robot. Such figures make progress more precisely measurable than a spectacular demonstration alone.
A secure grip needs the right force
The Learning Algorithms and Systems Laboratory at EPFL is also investigating how robotic hands handle objects. Its research goals include learning from human demonstrations and a stable grip whose force is adjusted during movement. Knowing the shape of an object is not enough for that: the grip must suit the specific task and remain secure.
In Switzerland, different parts of a possible supply chain are thus emerging: learning methods, controls and mechanical hands. At the ETH spin-off Mimic, use in manufacturing and logistics is the focus. EPFL also examines hands whose degrees of freedom differ from those of humans. For this work, the measure is not the most lifelike shape but the safe handling of objects. Which of these abilities a later household robot will take on will have to be judged by its published tests.
Sources
- Meta: research report on V-JEPA 2, 11 June 2025
- ETH Zurich: learning robotic hands and Mimic Robotics, 16 December 2025
- ETH Soft Robotics Lab: research project Fingers as Legs
- EPFL LASA: dexterous manipulation and grasping
- Reuters via MarketScreener: report on Meta and Fourier, 31 October 2026
- Header image: Jürg-Peter Hug, Wikimedia Commons
- Header image licence: CC BY-SA 3.0
- Article image: Sayanesy, Wikimedia Commons
- Article image licence: CC0 1.0
