After two years spent celebrating the race for giant language models, another revolution is taking hold without fanfare: that of autonomous robots.
Drones, industrial cobots, delivery vehicles, mobile defense platforms, service humanoids: everywhere, artificial intelligence is leaving the cloud to take shape in the physical world. The International Federation of Robotics is already talking about a “ChatGPT moment for physical AI.” But behind this promise lies a simple and largely underestimated constraint: a robot has neither the bandwidth, nor the battery, nor the tolerable latency to run a model with 70 billion parameters. Autonomous robotics will not be driven by ever-larger models. It will be done through smaller, faster, and above all embedded models.
AI that comes down from the cloud to the machine
The CETaS report (Alan Turing Institute, March 2026) identifies four major trajectories by 2035: foundation models for robotics, world models simulating the physical environment, versatile humanoids and swarms of small autonomous robots. These four trajectories share one characteristic: they are based on an AI which must execute at the edge, without depending on a stable link with a remote computing center. The global market for humanoids could reach $38 billion in 2035. That of service robots and cobots is already growing at double digits. None of these systems can afford a round trip to the cloud between perception and action.
Physical constraint redefines AI
A model useful to a robot must fit in a few gigabytes of memory, consume a few watts and respond in a few milliseconds. These three constraints — weight, energy, latency — have become the new frontier of AI engineering. They apply to the reconnaissance drone as well as to the robot vacuum cleaner, to the assembly arm as to the autonomous vehicle. The most striking example comes from NASA: the Perseverance rover now carries foundation models derived from next-generation vision architectures to reduce dependence on communication cycles with Earth. Deep space did not wait for hyperscale; he directly adopted compressed models.
Dual use as an industrial reality
Autonomous robotics is, by nature, a dual technology. Europol, in its Unmanned Future(s) report, recalls that the war in Ukraine has become a global accelerator: 1.5 million FPV drones produced in 2024, a target of 4.5 million for 2025, more than 200 domestic companies, 10,000 drones equipped with on-board AI acquired in a single year. The same components, the same software architectures and often the same supply chains irrigate civil uses: last mile logistics, infrastructure inspection, precision agriculture, industrial maintenance. On the defense side, swarms of small, low-cost robots; on the civilian side, fleets of collaborative cobots. In both cases, each unit carries little calculation. The value is shifting toward the ability to run compact, robust, distributed models.
The techniques that make small models possible
Several families of techniques converge to give these machines intelligence that can be used on site. Knowledge distillation transfers knowledge from a large model to a much smaller model. Quantization reduces the numerical precision of weights without significantly degrading quality. Compression by tensor networks, from computational physics, eliminates the internal redundancies of weight matrices and opens the way to reduction factors greater than those of classical approaches. Small specialized language models (edge SLMs) demonstrate that a model trained for a restricted domain can outperform, for the same task, a general model ten times larger. Finally, neuromorphic architectures, which players like Intel have been exploring for several years, promise energy efficiency of another order of magnitude, essential for long-term autonomous robots.
A sovereignty issue for Europe
The two reports converge on a harsh observation for the continent. CETaS ranks the UK 24th in the world for robotics adoption. Europol describes a Europe “critically dependent on foreign companies for advanced robotics hardware and software”. The critical chains are concentrated outside our borders: precision harmonic gearboxes dominated 80% by a single Japanese manufacturer, critical minerals controlled by China, robotics software platforms mainly developed across the Atlantic. Focusing on smaller, more efficient and locally deployable models is not only a technical choice; it is a credible path to regain autonomy on the AI layer, where the battle for hyperscale has already been lost.
Governance and security: the next project
Europol highlights a regulatory blind spot: the European frameworks on drones and autonomous systems were designed before the massive arrival of on-board AI. Neither the certification of the behavior of a compressed model, nor the traceability of a decision taken in a few milliseconds by a robot, nor liability in the event of an error currently have a solid legal basis. CETaS adds a concern: the scarcity of physical training data — manipulation sequences, multi-sensor interactions, realistic environments — constitutes a barrier as serious as hardware limitations. Without governance frameworks adapted to small embedded models, public trust could erode even before uses are deployed on a large scale.
The moment of choice
The trajectory is clear: by 2035, the artificial intelligence that will really count will not be that of giant computing centers, but that which fits in the on-board computer of a drone, in the arm of a cobot or in the motherboard of an autonomous vehicle. This AI is not obtained by stacking parameters; it is constructed through compression, distillation, quantification and specialization. For Europe, and for France in particular, this is a rare industrial opportunity: that of making efficiency, and not size, the hallmark of our robotics ecosystem. The bet on small models is not a fallback by default. This is the very condition for viable, sovereign autonomous robotics.
