Generative Bionics and AMD Plan European Physical AI Data Factory to Train Humanoid Robots
Italian robotics company Generative Bionics has announced plans to establish a large-scale physical artificial intelligence training infrastructure in Europe, combining real-world humanoid robot operations with advanced digital simulations powered by AMD computing technology. The initiative aims to address one of the most significant challenges facing intelligent robotics: developing machines capable of performing complex physical tasks reliably outside controlled laboratory environments.
The company plans to open its first 5,000-square-meter facility in Italy by the fourth quarter of 2027, describing it as the foundation of what it intends to become Europe's largest physical AI training network. The project will eventually expand to three locations, including another European facility and a third site serving international markets.
Unlike conventional AI data centers, which primarily process digital information, the proposed facilities will generate training data through physical interaction. Fleets of Generative Bionics' Gene.01 humanoid robots will operate within reconstructed industrial environments, collecting information about movement, manipulation, environmental conditions, and task execution.
The initiative reflects a broader transition in artificial intelligence from systems that interpret and generate digital content toward machines capable of interacting with the physical world. While large language models have demonstrated remarkable capabilities in text generation and reasoning, industrial robots must also understand spatial relationships, respond to unpredictable conditions, and execute actions safely.
By combining physical experimentation with large-scale simulation, Generative Bionics hopes to accelerate the development of humanoid robots that can operate effectively in manufacturing, logistics, and other demanding industrial environments.<Cite ref="turn467540view0"/>
Building a Data Factory for Physical Intelligence
The central concept behind the project is the creation of a continuous training environment where humanoid robots can learn from both real-world experiences and digitally generated scenarios.
At the Italian facility, Generative Bionics intends to reproduce industrial workplaces with equipment, materials, objects, and operational conditions similar to those encountered in actual factories. Fleets of Gene.01 robots will repeatedly perform physical tasks, generating data about their actions, sensor readings, interactions, and outcomes.
This information will support the development of World Action Models, AI systems intended to connect an understanding of the physical environment with the actions required to complete specific tasks.
A robot performing an industrial operation must do more than recognize an object. It needs to determine how to approach it, estimate the forces required for manipulation, account for obstacles, and adjust its movements when conditions change. These capabilities depend on training data that captures the complexity of physical interactions rather than relying exclusively on images or text.
Collecting such data presents substantial challenges. Real-world experiments require physical equipment, dedicated facilities, human supervision, and significant amounts of time. Robots may also encounter mechanical limitations or potentially hazardous situations during training.
Generative Bionics intends to address these constraints by combining physical training with digital twins, creating a development process in which real-world observations inform simulations and simulated improvements are subsequently validated using physical robots.
This approach could allow developers to explore many more operating conditions than would be practical through physical testing alone, while maintaining a connection between simulated performance and actual robotic behavior.
Digital Twins and the Real-to-Sim-to-Real Training Cycle
A major component of the proposed infrastructure is the use of digital twins: virtual representations of physical environments, equipment, and robotic systems.
Once robots perform tasks inside the training facility, the company plans to recreate those environments digitally. Developers can then modify variables such as object placement, materials, equipment configurations, environmental conditions, and potential failure scenarios.
For example, a humanoid robot trained to manipulate components on an assembly line could be tested virtually against thousands of variations in component position, weight, or surrounding obstacles. The resulting simulations would help researchers evaluate how well its control systems adapt when conditions differ from the original training environment.
The company describes this process as a real-to-sim-to-real cycle. Physical robots generate observations that improve digital models, simulations provide additional training and testing opportunities, and the resulting AI systems return to physical robots for validation.
The approach is designed to reduce the gap between simulation and reality, a longstanding challenge in robotics research. Simulated environments can accelerate development, but they cannot always reproduce real-world friction, material deformation, sensor imperfections, or unexpected physical interactions with sufficient accuracy.
Repeated validation against physical systems is therefore essential. A model that performs successfully in thousands of virtual scenarios may still require substantial refinement before it can operate safely and consistently in an actual industrial setting.
Generative Bionics believes that integrating physical experimentation and simulation within a dedicated infrastructure will make the training process more scalable. Rather than treating simulation as a separate development stage, the company intends to make it part of a continuous cycle of data generation, model improvement, and real-world testing.
AMD Helios Infrastructure to Power Large-Scale Robot Training
The project will rely on AMD computing infrastructure based on its Helios rack-scale architecture, which is designed to support demanding AI workloads through high-performance computing resources.
According to the companies, Helios systems will process data collected from physical robots, generate synthetic training information, support the development of physical AI models, and execute large numbers of digital twin simulations in parallel.
This computing capability is important because robotics training can involve several demanding workloads simultaneously. Developers may need to process high-volume sensor data, train machine learning models, simulate complex physical environments, and evaluate thousands of potential robot actions.
Running these workloads at scale requires substantial computational capacity, particularly when the goal is to train fleets of robots rather than individual experimental machines.
The partnership extends an existing relationship between AMD and Generative Bionics. AMD Ventures participated in the Italian company's €70 million seed funding round, while AMD technology is also used within the Gene.01 humanoid platform.
The collaboration therefore connects computing resources at two different levels: onboard processing that supports robot operation and centralized infrastructure that handles large-scale training and simulation.
This distinction reflects the emerging architecture of physical AI systems. Robots must make certain decisions locally to respond to their immediate surroundings, while computationally intensive model training and experimentation can take place within centralized infrastructure.
Daniele Pucci, CEO of Generative Bionics, has positioned the data factory as an essential step toward developing what the company calls an Action Intelligence layer for physical AI. The objective is to train robotic capabilities extensively before deploying them in customer environments, reducing the amount of experimentation required during real industrial operations.<Cite ref="turn467540view0"/>
Europe's Push for Industrial AI Sovereignty
Beyond the technical ambitions, the project carries broader implications for Europe's position in artificial intelligence and advanced manufacturing.
Much of the global AI infrastructure market is concentrated around a relatively small number of technology companies and computing platforms. Building dedicated physical AI training facilities within Europe could help regional robotics developers maintain greater control over their data, models, and development infrastructure.
Generative Bionics has emphasized that its planned network will support the development and operation of industrial humanoid systems within Europe. AMD has similarly framed the partnership as an opportunity to strengthen European industrial innovation and AI sovereignty.
The proposed three-site network suggests that the company is pursuing a distributed infrastructure strategy rather than concentrating all development activities within a single facility.
The first location in Italy will establish the initial physical training environment, while a second European facility and a third international site are intended to expand the network's reach. However, the company has not yet disclosed the precise locations, construction budgets, or operational schedules for the additional sites.
These details will be important in assessing the initiative's eventual scale. Generative Bionics describes its planned infrastructure as Europe's largest physical AI training facility, but that remains a company ambition rather than an independently established operational ranking.
The project's success will also depend on factors beyond computing power. Industrial humanoids must demonstrate reliability, energy efficiency, maintainability, and economic value before businesses can justify deploying them at scale.
Dedicated training infrastructure could help address some of these barriers, but commercial adoption will ultimately depend on whether the resulting robots can perform useful tasks consistently and at competitive costs.
Why Physical AI Data Factories Could Transform Robotics
The significance of Generative Bionics' announcement lies in its attempt to industrialize the process of training humanoid robots.
Traditional industrial automation has achieved considerable success by using specialized machines designed for repetitive tasks in structured environments. Humanoid robotics pursues a different objective: creating more adaptable systems capable of operating in workplaces originally designed for people.
Achieving that flexibility requires extensive training across different environments and physical situations. A robot that succeeds in one carefully configured factory may struggle when equipment layouts change, objects are positioned differently, or unexpected obstacles appear.
Physical AI data factories could provide a more systematic way to address these limitations by generating diverse training experiences and continuously validating improvements.
The model also introduces opportunities for more rigorous testing. Developers could evaluate robotic systems against repeatable scenarios, measure performance across environmental variations, and identify failure conditions before deploying machines into operational workplaces.
Nevertheless, large-scale simulation does not eliminate the need for real-world safety assessments. Physical robots introduce risks associated with moving machinery, unexpected behavior, and direct interaction with workers. Any deployment strategy must therefore incorporate independent validation, appropriate safeguards, and compliance with applicable industrial safety requirements.
The planned Italian facility is not expected to become operational until late 2027, and the company has not yet demonstrated that its proposed training infrastructure will deliver the performance improvements it anticipates. Its long-term impact will depend on execution, the quality of the collected data, and the ability to translate simulated learning into dependable real-world behavior.
Even with those uncertainties, the announcement represents an important direction for the robotics industry. As humanoid developers seek to move beyond prototypes and demonstrations, access to large-scale physical training data may become a critical competitive advantage.
Generative Bionics and AMD are betting that the next major advances in robotics will depend not only on better AI models or more capable hardware, but on the infrastructure required to teach machines how to operate in the physical world. If successful, their European data factory network could help establish a new foundation for training, testing, and deploying industrial humanoid robots at scale.