Insider Brief
- Generative Bionics plans to build a multi-site Physical AI Data Factory, beginning with a 5,000-square-meter hub in Italy expected to open by Q4 2027, to generate real-world and simulated data for training its GENE.01 industrial humanoid.
- The company said the facilities will combine physical robot operations with digital-twin simulation in a real-to-sim-to-real training loop, using data from vision, motion, force, proximity, temperature, human interaction and the robot’s full-body multimodal skin.
- AMD will serve as the AI infrastructure partner, with Generative Bionics planning to use Helios rack-scale systems to process physical and synthetic data, train World Action Models and run large numbers of digital-twin simulations in parallel.
Italian startup Generative Bionics announced it plans to build a multi-site Physical AI Data Factory designed to generate the real-world and simulated data needed to train humanoid robots for industrial use.
According to Generative Bionics, the first 5,000-square-meter hub, expected to become Europe’s largest physical AI training facility for humanoid robots, is planned for Italy to begin operating by the fourth quarter of 2027. A second facility would follow elsewhere in Europe, with a third serving global markets.
The company plans, announced at CES Unveiled Milan, to use the facilities to reproduce industrial environments where fleets of its GENE.01 humanoid robots can perform tasks continuously. Those operations will generate physical data that can then be combined with large-scale digital-twin simulations to train and test the robots before they are deployed at customer sites.
The company is building the network around a problem facing humanoid developers in that unlike large language models, which can train on vast amounts of online information, robots need data describing how physical tasks unfold in the real world.
For GENE.01, that includes vision and motion data along with information about contact, force, proximity, temperature and human interaction. Signals will also come from the robot’s full-body multimodal skin, the company noted.
Real-to-Sim-to-Real Training
Each physical environment and robot configuration will also be recreated as a digital twin. The same task can then be repeated across thousands of simulated variations involving changes in layouts, objects, materials, operating conditions and failure modes.
Models trained or tested in simulation can be sent back to physical robots for validation, while data collected from those robots feeds subsequent training cycles.
Generative Bionics said this real-to-sim-to-real process is intended to let it configure and test GENE.01 for specific industrial settings before deployment. Data from robots already operating in the field would later be added to future training cycles.
The company said it plans to use the facilities to train World Action Models and test the broader robot system, including perception, control, safety behavior and task execution.
AMD Infrastructure
AMD will serve as the AI infrastructure partner for the Data Factory network, according to Generative Bionics. The company plans to use AMD Helios rack-scale systems for processing physical and synthetic data, training physical AI models and running digital-twin simulations in parallel.
Generative Bionics said it intends to locate robotics, AI and data-engineering teams alongside the computing infrastructure so they can manage the pipeline from raw robot data through dataset preparation, model training, validation and deployment.
“With AMD at our side since day one, on board GENE.01 and now at the heart of the Data Factory, we are building this network here in Europe, so that the data, the models and the infrastructure behind our industrial humanoids are developed and operated on the continent,” Generative Bionics CEO Daniele Pucci said.
