NEURA Robotics and RWTH Aachen University are establishing a 3,000m² physical artificial intelligence training facility where industrial and humanoid robots will learn, test, and validate practical tasks before entering production environments.
Located at RWTH Aachen’s Hightech Campus Melaten, the NEURA Gym is scheduled to open towards the end of 2026. Around 20 university institutes are expected to participate, alongside manufacturers and technology companies developing applications through the facility.
Physical robot operation will be combined with simulation so that machines can perform tasks in controlled environments, generate training data, and improve their behaviour before deployment. Engineers will also be able to investigate failure modes without interrupting an operating factory.
Initial applications are expected across manufacturing, automotive production, medical technology, and the circular economy, where many processes remain difficult to automate through conventional fixed programming because components, materials, positions, and surrounding conditions vary.
Traditional industrial robot cells perform most consistently when each component arrives in a predictable location and the required movement can be defined in advance. Physical AI is intended to give machines greater capacity to interpret their surroundings, plan actions, compensate for variation, and improve through repeated operation.
David Reger, founder and chief executive officer of NEURA Robotics, said: “The biggest bottleneck in Physical AI is no longer intelligence, it’s experience.”
By providing that experience away from customer production, the Aachen facility will allow engineers to recreate working environments, collect operational data, examine failures, and refine applications before downtime or unsafe behaviour can disrupt a live process.
NEURA says ten gyms are under development and expects five to be operating by the end of 2026 across Europe, the United States, and China. The company has linked the programme to a funding round of up to $1.4 billion intended to support industrial development of physical AI.
Robot training becomes industrial infrastructure
Many industrial tasks are not documented in a form that can be used directly for machine learning, even where experienced employees perform them reliably. Operators may compensate for a misaligned component, flexible material, worn tool, damaged surface, or changing process condition without recording each decision explicitly.
Capturing that knowledge requires more than filming the work because a training system may need information from force sensors, cameras, joint positions, tooling, process equipment, inspection results, and the condition of the final product.
The data must distinguish acceptable variation from genuine failure, while identifying which action produced a successful result. Poor examples or incorrectly labelled outcomes can teach a robot to repeat behaviour that appears effective during training but creates quality problems elsewhere.
Simulation can increase the number of examples and allow hazardous scenarios to be explored, although it cannot reproduce friction, wear, lighting, vibration, compliance, contamination, and unexpected human activity perfectly. Physical testing remains necessary to close the gap between software performance and dependable operation around real machinery.
The gym model may alter how automation projects are developed and purchased. Instead of specifying a fixed cell around one task, manufacturers could train a more adaptable platform against several processes before deployment, creating scope for reuse when products or factory layouts change.
Greater adaptability also increases dependence on software maintenance and continuing validation. A robot that has learned a process cannot be assumed to perform indefinitely when components, suppliers, lighting, packaging, tooling, or surrounding equipment change.
Safety systems must remain predictable even when the robot’s task behaviour becomes more flexible. Safe speed and force limits, collision detection, emergency stopping, access controls, operating boundaries, and software update procedures will require validation around the complete application.
Industrial data ownership introduces another layer of control because factory processes may reveal proprietary information about materials, tolerances, cycle times, tooling, and quality methods. NEURA says participating companies will retain control of their data, but storage, separation, access, and reuse must be defined in practice.
Monitoring for model drift is likely to become part of routine maintenance alongside mechanical wear, lubrication, calibration, and sensor condition. Performance that changes gradually may be harder to detect than a conventional component failure, particularly where the robot continues operating while producing less consistent results.
RWTH Aachen contributes expertise spanning production engineering, machine tools, computer science, control, materials, and human factors. Physical AI sits between those disciplines, since useful intelligence must operate within the constraints of machinery, process physics, employee safety, and production economics.
The facility’s industrial performance will be measured through applications that move from training into sustained production with acceptable uptime, cycle time, quality, and maintenance cost. Humanoid demonstrations can attract attention, but repeatable factory operation will determine whether physical AI becomes a practical automation platform.
