AI and Open Source: Accelerating the Arrival of Artificial Intelligence in the Physical World
For several years now, our perception of artificial intelligence has been strongly shaped by generative models. We ask it to write texts, analyze documents, produce images, search for information, or assist us with various intellectual tasks. The interaction takes place essentially in a digital universe and, when an error occurs, its consequences generally remain confined to that environment.
Another development is nonetheless unfolding in parallel. Artificial intelligence is beginning to move beyond screens to interact directly with the physical world. Robots are learning to recognize objects, understand their environment, plan their movements, handle various parts, and adapt their motions when the situation changes. Autonomous vehicles, industrial robots, exoskeletons, drones, and other intelligent machines are thus beginning to bring together perception, reasoning, and action. This evolution is increasingly referred to as physical AI, or Physical AI.
Making an artificial intelligence work in a digital environment is already complex. Enabling it to act reliably in an unpredictable physical world represents a challenge of an entirely different magnitude. The European consortium MAESTRO illustrates a particularly interesting approach in this respect: bringing research and industry together around an Open Source software platform that provides some of the fundamental building blocks needed for intelligent robotics.
This encounter between physical AI and Open Source could play an important role in the evolution of robotics. To build tomorrow's intelligent machines, industry will need high-performing artificial intelligence models, but also software foundations that are robust, transparent, reusable, and scalable. The ability to share some of these foundations could allow researchers and industrial players to focus more of their effort on the innovations that genuinely differentiate their products.
The fundamental difference between generative artificial intelligence and physical AI lies in the relationship each maintains with its environment. A conversational model generally receives information and produces information. We send it a text, an image, or a document, and it returns an answer, an analysis, or new content. Physical AI must go through several additional steps. It must perceive its environment, understand enough of what is happening there, determine an appropriate action, and then translate that decision into movements or commands capable of producing a result in the real world.
An apparently simple instruction such as “pick up that box and place it on that pallet” quickly illustrates the complexity of the problem. A robot must identify the right box, determine its position, know its own configuration, plan the movement of its arm, avoid obstacles, grip the object with appropriate force, maintain its stability, and adapt its motion when reality differs from what was expected. Gravity, mass, collisions, friction, and unforeseen movements are all part of the problem. An obstacle may appear, a person may cross the planned path, or an object may sit a few centimeters farther away than expected. Unlike a purely digital environment, the physical world has no “Undo” button.
Industry has obviously been using robots for several decades, and some perform their tasks with remarkable precision and speed. Their effectiveness, however, often rests on tightly controlled environments. A part always arrives at the same place and in the same orientation, the robot repeats a programmed motion, and safety devices limit human access to its workspace. This predictability enables extremely high-performing automation. The challenge changes when the environment becomes variable, when objects present themselves differently, when several tasks must be performed, or when robots and humans share the same space directly.
Physical AI seeks precisely to increase this capacity for adaptation. The robot must progressively be able to perceive variations in its environment and adjust its behavior accordingly. This requires bringing together several disciplines. Perception serves to interpret the environment, physical models represent the robot and its interactions with its surroundings, planning algorithms determine how to reach an objective, optimization methods search for efficient trajectories, and control systems ultimately translate those decisions into actions the motors can execute. Artificial intelligence can intervene at various stages while remaining subject to the physical constraints of the system.
That is why a modern robot must also be seen as an immense software stack. Motors, articulated arms, wheels, legs, cameras, and sensors make up the visible part of the machine, but much of its intelligence remains invisible. Software mathematically represents its structure, computes its movements, detects collisions, simulates physical interactions, optimizes trajectories, and controls actions. MAESTRO aims precisely to bring several of these functions together into a coherent stack that includes components dedicated to robot dynamics, collision detection, optimal control, numerical optimization, and differentiable physical simulation.
This is where Open Source can become a particularly powerful accelerator. Its strategic value extends far beyond the simple fact that software is free. When a fundamental software infrastructure is open, different organizations can use the same base, examine it, test it, improve it, and develop their own innovations on top of it. Modern computing was largely built this way. A company developing a web application generally does not recreate its own operating system, its programming language, its web server, or all of its databases. It builds on technology layers accumulated over years by communities, companies, and research organizations.
This collective accumulation allows each new generation of developers to start further along than the previous one. Robotics could benefit from the same phenomenon. If researchers and industrial players have common foundations for modeling, simulation, optimization, and control, they can devote more resources to the problems specific to their products. An automaker and an exoskeleton manufacturer can use some of the same fundamental libraries while developing radically different products. An aerospace manufacturer and an industrial robotics company can share certain mathematical tools while retaining their data, their models, their processes, their interfaces, and their know-how.
Sharing foundations therefore does not mean sharing competitive advantage. Linux, Python, Kubernetes, and several other Open Source technologies are already used by directly competing companies without making their products uniform. Differentiation is built in the way those foundations are combined, adapted, and enriched by the knowledge specific to each organization. In the field of physical AI, this pooling can be particularly important, since the fundamental problems of robotics are complex enough to absorb immense resources before a company even begins working on what makes its product truly distinctive.
Open Source can also help reduce the distance between scientific research and industry. Developing a high-performing algorithm in a laboratory is an important step, but its daily use in a factory demands far more. It must be turned into software that is robust, documented, maintainable, compatible with other components, and performant enough to run continuously. That transition requires substantial engineering work. A well-structured Open Source ecosystem can serve as a bridge between these two worlds: researchers can integrate their advances into a common platform, industrial players can confront them with real problems, and the difficulties observed in production can then shape research priorities.
MAESTRO was designed precisely around this relationship. Scientific organizations such as Inria and the CNRS work there alongside industrial partners from various sectors. Research feeds the technological foundations while companies confront those technologies with industrial realities. This loop between research, engineering, and real-world use can help accelerate the maturation of technologies that would otherwise remain confined to laboratories for longer.
Openness of software also has a particularly important dimension when that software takes part in controlling a physical machine: auditability. When software is involved in the movements of an industrial robot, an exoskeleton, or a system operating in a sensitive environment, the ability to understand how it works takes on a different value. Access to the code allows specialists to examine the components, understand the architecture, look for certain vulnerabilities, and verify their behavior. Open Source software is obviously not automatically secure. It may contain vulnerabilities, be poorly maintained, or depend on problematic components. Openness nonetheless creates an essential possibility: that of inspection.
This capacity can also contribute to technological sovereignty. Robotics and artificial intelligence are progressively becoming strategic industrial capabilities. An industry entirely dependent on a software stack controlled by a handful of vendors may see its room to maneuver shrink when the licenses, prices, commercial directions, or availability of those technologies change. Open foundations can preserve more options, provided organizations also retain the skills needed to understand, maintain, and evolve them. Owning the code without owning the knowledge required to exploit it creates only theoretical sovereignty.
True technological autonomy therefore rests on a broader set of elements including access to technologies, skills, documentation, the vitality of communities, integration capability, and governance. An abandoned or incomprehensible Open Source stack can become a dependency as problematic as a proprietary technology. MAESTRO's objective of contributing to an open European software foundation for robotics thus takes on a dimension that goes beyond software development: it contributes to building a collective capacity to understand, master, and evolve certain fundamental technologies.
Common foundations can also reduce fragmentation. As robots multiply, each manufacturer could be tempted to develop its own tools, interfaces, formats, and methods. That diversity can stimulate innovation, but it can also make integration extremely costly. Common components can progressively create a shared technical language, make it easier to adapt an algorithm across different platforms, allow researchers to compare their approaches, and make skills more transferable between companies.
The history of the Web provides an interesting precedent. Its growth did not come from a single company controlling the entire infrastructure, but largely from common standards allowing technologies developed by different players to communicate with one another. Physical AI may likewise need sufficiently common foundations and interfaces to reach genuine industrial scale. The objective is not to make robots uniform, but to reduce the amount of work devoted to continually rebuilding the same fundamental layers.
This approach nonetheless requires solid governance. Publishing code on the Internet is not enough to create a durable industrial infrastructure. Vulnerabilities must be fixed, new versions tested, changes documented, and contributions evaluated. Technological directions must be coordinated and compatibility must be managed over time. These responsibilities become particularly important when software is integrated into critical industrial environments. Open Source then becomes a collective infrastructure, supported by an ecosystem capable of ensuring its continuity, rather than a mere publicly accessible code repository.
As these foundations and artificial intelligence capabilities advance, they could also shift the economic boundary of automation. Traditional industrial robotics excels in repetitive, predictable, and highly standardized environments. That characteristic partly explains why certain forms of automation are very profitable for large production volumes and far less accessible to companies manufacturing small runs or many different products. Robotics that is better able to perceive and adapt could make automation relevant in far more variable environments.
A production line could thus accommodate more variation without requiring complete reprogramming. Robots could learn new manipulations, mobile machines could operate in environments shared with people, and some systems could adapt their movements based on what they observe. This flexibility could progressively bring robotization within reach of companies that until now had neither the volume nor the level of standardization needed to justify certain forms of automation.
This evolution does not mean industrial environments will simply become spaces emptied of their workers. A significant part of these developments concerns precisely the ability of robots to work better in environments where people are present. Adaptive systems can assist employees with strenuous handling, exoskeletons can augment or restore certain physical capabilities, and mobile robots can take charge of moving materials while people focus on tasks requiring more judgment, experience, human connection, or dexterity.
Skills will nonetheless have to evolve. Organizations will need people capable of configuring, supervising, maintaining, and improving these systems. Operators will need to understand enough about how they work to recognize unusual behaviors and intervene when conditions fall outside the expected range. As we are already seeing with generative AI and AI agents, technological transformation therefore also becomes a transformation of processes, roles, and organizational knowledge.
The arrival of artificial intelligence in the physical world finally brings together two disciplines that organizations have sometimes treated separately: cybersecurity and operational safety. When a computer system controls a machine, a digital compromise can produce a physical consequence. Changing the parameters of an application and changing the trajectory of a robotic arm obviously do not present the same level of risk. The identity of machines, their permissions, the integrity of software, update mechanisms, the security of communications, and the provenance of software components therefore become elements directly tied to the safety of operations.
The principles of least privilege, Zero Trust, segmentation, and defense in depth will thus have to extend to robotic infrastructures. Open Source can contribute to the auditability of certain components, but security will always depend on the complete architecture in which they are integrated. The ability to examine a library is not enough if identities are poorly managed, if permissions are excessive, if updates are not controlled, or if communications between systems can be compromised.
This convergence represents an important evolution of Hypersecurity. As artificial intelligence acquires the capacity to act in the physical world, protecting digital systems becomes directly tied to protecting physical operations. Security must then cover the entire chain that links perception, data, models, software, identities, permissions, decisions, and the actions carried out by the machine. The further artificial intelligence enters the physical world, the more cybersecurity itself becomes a question of operational safety.
Physical AI represents a major evolution of artificial intelligence. It adds to information processing the ability to perceive an environment, interpret certain situations, and turn digital decisions into actions in the real world. This evolution demands far more than high-performing models. Robots must understand their own mechanics, anticipate their movements, avoid collisions, optimize their trajectories, manage their interactions with objects and people, and react quickly enough when their environment changes. Behind each of these behaviors lies a considerable software infrastructure.
The Open Source approach carried by initiatives such as MAESTRO becomes particularly interesting in this context. By pooling certain fundamental building blocks of modeling, simulation, optimization, and control, researchers and industrial players can build on a common base, examine it, contribute to improving it, and focus more of their resources on the innovations specific to their applications. This approach can also accelerate the transition between research and industry, improve interoperability, foster auditability, and preserve a collective capacity to understand and evolve technologies that could become essential to tomorrow's industry.
For Quantum Beyond, this evolution connects several dimensions we see as deeply linked: artificial intelligence, Edge infrastructures, machine identity, Hypersecurity, technological sovereignty, governance, and organizational transformation. As digital systems acquire the capacity to act in the physical world, organizations will need to understand the architectures that control them, secure their interactions, and retain enough technological mastery to evolve them. Our experts can work alongside technology, operations, and cybersecurity teams to structure this transition and connect these new capabilities to the infrastructures and skills already present in the organization.
Open Source will obviously not solve the challenges of physical AI on its own. Its contribution can nonetheless be decisive if it allows companies, laboratories, and manufacturers to collectively build certain foundations rather than continually devoting their resources to recreating them separately.
Generative AI showed us what becomes possible when machines learn to handle language and information. Physical AI now opens a far more concrete stage: teaching those machines to understand our world well enough to act in it. If part of the necessary foundations remains open, masterable, and shared, we could also learn to build this new generation of machines far more quickly together than separately.