Robotics: The Future Rests on Compact AI Models

A Quiet Revolution in Autonomous Robotics

While giant language models have dominated recent discussions, a new revolution is quietly emerging in the field of autonomous robots. These machines, ranging from drones to service humanoids, delivery vehicles, and mobile defense platforms, are now integrating artificial intelligence directly into their hardware, without relying on remote computing centers. The International Federation of Robotics refers to a "ChatGPT moment" for physical AI, highlighting the growing importance of this transition. However, robots are limited by physical constraints such as bandwidth, battery life, and latency, making giant models unsuitable. The future of autonomous robotics therefore relies on smaller, faster, and embedded models.

Embedded AI: A Major Challenge for 2035

According to a report from CETaS published by the Alan Turing Institute in March 2026, four major trajectories are emerging for robotics by 2035: foundation models, models simulating the physical environment, versatile humanoids, and swarms of small autonomous robots. These trajectories share a key characteristic: AI must operate locally, without dependence on a cloud connection. The humanoid market could reach $38 billion by 2035, while the service robots and cobots market is already experiencing double-digit growth. These systems cannot afford back-and-forth trips to the cloud between perception and action.

Physical Constraints Redefining AI

To be useful, an embedded AI model must be lightweight, energy-efficient, and responsive. These constraints of weight, energy, and latency are now at the heart of AI engineering. An effective model must fit within a few gigabytes of memory, consume a few watts, and respond within milliseconds. Concrete examples, such as NASA's Perseverance rover, illustrate this trend. The rover uses compressed foundation models to reduce its dependence on communications with Earth, demonstrating the importance of compact models in extreme environments.

The Duality of Autonomous Robotics

Autonomous robotics is intrinsically dual, with both civilian and military applications. Europol's report, "Unmanned Future(s)," highlights that the war in Ukraine has accelerated this trend. In 2024, 1.5 million FPV drones were produced, with a target of 4.5 million for 2025. Over 200 domestic companies and 10,000 drones equipped with embedded AI were acquired in just one year. These same technologies are being used in civilian applications such as logistics, infrastructure inspection, and precision agriculture.

Techniques for Smaller AI Models

Several techniques enable the creation of efficient AI models for autonomous robots. Knowledge distillation transfers knowledge from a large model to a smaller one. Quantization reduces numerical precision without degrading quality. Tensor network compression eliminates internal redundancies, and specialized language models demonstrate their effectiveness in restricted domains. Small specialized language models (edge SLMs) show that a model trained for a specific domain can outperform a general model ten times larger on the same task. Neuromorphic architectures, explored by companies like Intel, promise increased energy efficiency, which is essential for autonomous robots.

A Sovereignty Challenge for Europe

Reports highlight Europe's critical dependence on foreign companies for advanced robotic technologies. CETaS ranks the UK 24th globally for robotic adoption, while Europol emphasizes foreign dominance over critical supply chains. Precision harmonic reducers are 80% dominated by a single Japanese manufacturer, and critical minerals are controlled by China. Moreover, robotic software platforms are predominantly developed across the Atlantic. Critical supply chains are concentrated outside our borders. Betting on smaller, more efficient AI models could allow Europe to regain technological autonomy.

Governance and Security: Challenges Ahead

Europol points out a lack of an appropriate regulatory framework for autonomous systems equipped with embedded AI. European frameworks on drones and autonomous systems were designed before the massive arrival of embedded AI. Certifications, decision traceability, and accountability in case of errors require solid legal foundations. CETaS also raises a concern: the scarcity of physical training data—manipulation sequences, multi-sensor interactions, realistic environments—constitutes a bottleneck as serious as hardware limitations. Without adequate governance, public trust could be compromised even before these technologies are widely adopted.

Towards Embedded AI by 2035

By 2035, the AI that will matter will not be that of giant computing centers, but that integrated into autonomous robots. This AI will develop through techniques of compression, distillation, and specialization. For Europe, this is an industrial opportunity to stand out by focusing on efficiency rather than size, and to strengthen its technological sovereignty.