TUM and NEURA Open a Robot Gym for the Skills Factories Struggle to Teach
The Munich facility brings human demonstrations, robot practice and simulation into a shared research and training operation.
Published October 27, 2026Intelligence / Robotics / Humanoid Robotics
Folding a box or fitting two parts together can look simple until someone has to teach a machine how to do it. The Technical University of Munich (TUM) and NEURA Robotics opened a facility on October 22 to tackle that problem through repeated physical practice. Their RoboGym at LabCampus near Munich Airport occupies more than 2,300 square meters. NEURA says industrial partners had already been training workflows there since September, making the opening a milestone for an operation under way. [1]
TUM describes 48 training cells with about 100 humanoid and service robots. The university's initial focus is industrial work and intralogistics, including folding boxes, assembling components and using tools. Many machines can generate experience across different tasks and conditions. Researchers can use the center to study learning and reliability systematically, while companies can explore applications before attempting a wider deployment in their own facilities. [2] [3]
The training combines three sources of experience. Robots collect sensor data as they move, make contact and complete actions. People demonstrate workflows through motion capture, sensor-equipped suits and teleoperation. Simulation then introduces additional variations for training and testing on physical machines. NEURA says the center uses these methods together. Their practical purpose is to connect the appearance of a task with the forces, timing and corrective movements required to carry it out. [1]
Much of that knowledge resides with the people who perform the work. An experienced operator may feel when a part starts to bind or change a grip before a flexible object slips. Capturing such adjustments requires more than a video of the final successful motion. A useful training program needs examples of variation and recovery, together with a way to identify which action produced a good result. The facility's access to real industrial tasks offers a setting in which to investigate those details.
TUM says the resulting data feeds a curated cloud data lake for developing physical artificial intelligence (AI) models. Its detailed account also describes shared software for coordinating robots, annotating data and improving models. That common infrastructure could reduce the effort needed to organize individual experiments and make results easier to compare. The larger research goal is transfer: a skill learned under one set of conditions should help a robot cope with a new object, task or machine. TUM presents that capability as an objective of the program. [2] [3]
NEURA used the opening to introduce its 4NE1 Gen 4 humanoid, describing a design intended for series production. The training center also accommodates robot arms and other mobile systems. That mix is useful because industrial tasks impose different requirements on reach, mobility and precision. The evidence so far establishes a facility, participating companies and a training approach. A claim about factory productivity would additionally need measured task completion, intervention rates and performance after a skill moves into a customer's working environment. [1] [3]
The most revealing results will show how much new experience a robot needs when the job changes. Researchers could compare training time and failure rates for a familiar task, a variation and an unfamiliar setting, while recording the human effort behind each result. Those measures would connect the center's substantial infrastructure to the problem it was built to solve. RoboGym gives human process knowledge a place in robot development; the next stage is showing how reliably that knowledge travels beyond the training cell.