{"id":2066863,"date":"2026-08-04T13:21:49","date_gmt":"2026-08-04T11:21:49","guid":{"rendered":"https:\/\/www.luxsure.fr\/2026\/08\/04\/humanoids-can-europe-still-build-robot-brains\/"},"modified":"2026-08-04T13:24:43","modified_gmt":"2026-08-04T11:24:43","slug":"humanoids-can-europe-still-build-robot-brains","status":"publish","type":"post","link":"https:\/\/www.luxsure.fr\/en\/2026\/08\/04\/humanoids-can-europe-still-build-robot-brains\/","title":{"rendered":"Humanoids: Can Europe Still Build Robot Brains?"},"content":{"rendered":"\n

The next battle in artificial intelligence won\u2019t be fought solely in the realm of language models. It will shift to warehouses, factories, and repetitive tasks\u2014places where machines will have to learn to act without being guided step by step. <\/p>\n\n

There is a theatrical allure to the humanoid. A figure walking, a hand grasping a bottle, a mechanical torso crossing a room: the image alone is often enough to create the event. But the serious question is not whether the robot resembles a human. It is who programs its intelligence. <\/p>\n\n

Flexion Robotics, a Zurich-based company, fits precisely into this category. Not in the manufacturing of the body, but in that of the \u201cbrain\u201d: a software layer capable of transforming an instruction\u2014such as tidying up a room, moving a box, or handling an object\u2014into movements, decisions, and motor adjustments. In this architecture, the robot is no longer just a machine equipped with sensors; it becomes a physical agent, capable of learning a task in simulation before transferring it to the real world. <\/p>\n\n

This nuance is crucial. The humanoid robot market is currently saturated with brief demonstrations\u2014often impressive, but rarely conclusive. A video lasting just a few seconds reveals nothing about the number of engineers involved, the level of teleoperation, or the behind-the-scenes work required to produce a seemingly autonomous movement. True autonomy is evident elsewhere: in the duration of the task, in the variability of the environment, and in the ability to act without every movement having been preprogrammed by humans. <\/p>\n\n

This is where Flexion advocates a more low-key but more structured approach: training robots in simulation using reinforcement learning. Rather than having human operators\u2014equipped with helmets or protective suits\u2014repeat a task over and over, the company creates virtual worlds where the robot can fail millions of times without breaking any equipment, without tying up a team, and without limiting its learning to a single configuration. A table, a bottle, a light, an angle, a background: everything can be recombined in infinite ways. <\/p>\n\n

Physical data: the big missing piece<\/h2>\n\n

Large language models have been able to draw on the vast corpus of text available online. Robotics, on the other hand, lacks a natural equivalent. Human gestures have not been widely archived in a usable format. Grasping, pushing, screwing, carrying, avoiding, balancing: these actions occur everywhere, but data on them is not available in the same way as web pages. <\/p>\n\n

Teleoperation therefore emerged as a logical solution. However, it remains slow, costly, and imperfect. A human operating a robot remotely cannot feel the forces acting on the object. They can see, but they cannot truly touch it. The result is often imprecise data\u2014useful for getting started, but insufficient for broader application. <\/p>\n\n

Reinforcement learning turns the logic on its head. The robot no longer simply copies humans; it explores, fails, adjusts, and tries again. In simulation, this exploration becomes economically feasible. The cost of error decreases, the variety of situations increases, and data is no longer simply collected\u2014it is generated. <\/p>\n\n

The Details<\/strong>
A simple task, such as picking up a bottle in a fixed environment, can be learned with just a few dozen human demonstrations. But as soon as the lighting changes, the object moves, or the setting varies, the model loses its robustness. Simulation makes it possible to create thousands of variations of the same gesture without moving the robot from one location to another. <\/p>\n\n

The real issue: Who controls the movement?<\/h2>\n\n

Europe talks a lot about sovereignty in the cloud, models, and data. It still talks too little about sovereignty in machines capable of taking action. Yet an autonomous robot is not just abstract software. It moves around a warehouse, carries loads, handles objects, and interacts with human infrastructure. <\/p>\n\n

In this context, software dependency becomes an even more significant issue than in generative AI. If the hardware is European but the software \u201cbrain\u201d comes from elsewhere, control remains only partial. The \u201cbody\u201d belongs to the manufacturer; the \u201cintention\u201d belongs to the provider of the autonomy system. <\/p>\n\n

This may be where Europe still has an opportunity. It hasn\u2019t always dominated consumer platforms, but it possesses considerable industrial strengths: the automotive industry, machine tools, logistics, specialized robotics, and complex supply chains. Companies capable of large-scale production already exist. What\u2019s missing is the general-purpose software that would allow these machines to move beyond programmed actions and achieve adaptive autonomy. <\/p>\n\n

Flexion is therefore banking on a separation of roles. Manufacturers will handle the production of the hardware, while autonomy specialists will focus on the \u201cgrammar\u201d of movement. This choice is not merely technical; it is strategic. It is difficult for a startup to produce millions of machines. It can, however, equip various types of robots, learn from how they are used, and improve its model more quickly if its software runs on multiple hardware platforms. <\/p>\n\n

The Humanoid, or the Temporary Compromise<\/h2>\n\n

The term \u201chumanoid\u201d is fascinating because it promises a direct substitution: replacing human labor with a human-shaped machine. In reality, the human form is merely a starting point. In a warehouse with a flat floor, wheels may be preferable to legs. For carrying heavy loads, a body more like a gorilla\u2019s than a human\u2019s may be more suitable. The goal is not to copy human anatomy, but to access spaces designed for humans. <\/p>\n\n

That is why the first truly cost-effective applications will likely not be in the home. A robot that folds laundry in an apartment remains an appealing fantasy, but the home environment is too varied, too unpredictable, and too cluttered with unique objects. Industry and logistics offer a more predictable environment: cardboard boxes, standardized bins, order picking, packing, material handling, and repetitive movements. <\/p>\n\n

The first useful tasks could emerge quickly. But usefulness alone will not be enough. The economic threshold will be crossed when robots no longer need engineers behind the scenes to ensure the safety of every deployment. As long as autonomy requires constant human oversight, the promise of increased productivity remains unfulfilled. <\/p>\n\n

Humanoid robotics is therefore not yet at its spectacular stage. It is currently in its infrastructure phase. The body attracts attention, but the brain will determine its value. And for Europe, the question is no longer whether it will participate in the race. It is whether it will agree to build the bodies while others write the code. <\/p>\n\n

The next hand to move a crate in a European warehouse might have five fingers, two arms, or wheels. The details matter less than the invisible command that will direct its movement. <\/p>\n\n

\"\"<\/figure>\n","protected":false},"excerpt":{"rendered":"

The next battle in artificial intelligence won\u2019t be fought solely in the realm of language models. It will shift to warehouses, factories, and repetitive tasks\u2014places where machines will have to…<\/p>\n","protected":false},"author":2,"featured_media":2063666,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_lmt_disableupdate":"","_lmt_disable":"","footnotes":""},"categories":[51688],"tags":[78440,78443,51736,78441,78442],"class_list":["post-2066863","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-luxe-ia","tag-flexion-robotics","tag-humanoides-industriels","tag-ia-physique","tag-apprentissage-par-renforcement","tag-robotique-souveraine"],"_links":{"self":[{"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/posts\/2066863","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/comments?post=2066863"}],"version-history":[{"count":0,"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/posts\/2066863\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/media\/2063666"}],"wp:attachment":[{"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/media?parent=2066863"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/categories?post=2066863"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/tags?post=2066863"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}