Most robot videos are edited. A cut hides the failed grasp, a speed-up hides the hesitation, and the best take makes it online. At NVIDIA GTC Berlin 2026, held from October 20 to 22 at Station-Berlin, we did the opposite. At Booth 3031 in Hall 3, live in the event and online, our humanoid robot 4NE1 Wheels autonomously completed a full laundry cycle, again and again, continuously throughout the show.
The live demo: an eleven-minute laundry cycle without human intervention
The task looks simple because people do it every week. For a robot, it is a long chain of steps where every step depends on the one before:
- Sorting. 4NE1 separated light and dark garments.
- Washing. It loaded the washing machine and closed the door.
- Transferring. Once the cycle finished, it opened the door, unloaded the garments, moved them into the dryer and closed the dryer door.
The full sequence took around eleven minutes and ran without human intervention. To keep it watchable, the wash and dry cycles of the BSH appliances were shortened to about ten seconds. All other conditions were real life conditions for the robot, acting autonomously through our multimodal foundation model. On Wednesday, the demo was also streamed live on YouTube and LinkedIn.
The team made one deliberate trade-off. For this first public run, reliability and safety around visitors came before speed. The goal was to finish every cycle end to end. Speed will be tuned in later model updates.
Why a laundry cycle is hard for a robot: memory and senses beyond vision
The difficult moment comes when the machine door closes. The garments disappear from view, yet the robot still has to know where they are, what it already did and what comes next. Humans learn this, called object permanence, in their first two years of life. A robot needs it built in.
4NE1 runs on Neuron AI, the multimodal foundation model NEURA introduced on October 14. Its Perception-Language-Action (PLA) architecture brings 3D vision, touch, force, audio, language and the robot’s own state together with memory, and turns them into one stream of action. At the booth, that meant:
- Vision at two scales. A chest camera mapped the workspace, while cameras in the hands guided each grasp.
- Touch where cameras fail. Inside the drum, sight is blocked, so tactile sensing takes over.
- Sound as confirmation. A door latching or a connector snapping into place can be heard before it can be seen.
The model learns which sense to rely on at each stage of the task. The same principle carries into industry: in welding, for example, the quality of a seam can be judged by its sound.
The keynote: From Europe, for the World
On Thursday, Arne Nordmann, VP Technology & Innovation at NEURA Robotics, gave the talk “From Europe, for the World: Building the Ecosystem for Physical AI” at the Sponsored Expo Theater. His point: a capable robot is the visible end of an ecosystem that has to exist behind it.
He described how that ecosystem works in practice. In NEURA Gyms, domain experts such as master welders and manufacturing engineers physically teach robots their tasks. NVIDIA simulation tools then multiply that real-world data across hundreds of variations in lighting, object geometry and environment. Training runs in the cloud every day, while inference happens on the robot itself, on the compute built into its chest. Through the Neuraverse, a skill learned by one robot can be deployed across a whole fleet of different robot types, and partner companies can turn their own process knowledge into skills.
Data partnerships with industrial companies, including Schaeffler and Bosch, add real manufacturing processes to that pipeline. As Arne Nordmann put it, the result is German hardware and a German foundation model, operating German appliances, developed in Europe for deployment worldwide.
GTC Berlin was the perfect stage to show what Physical AI can already do today: a humanoid robot completing a complex, real-world task on its own, live and without cuts. I am grateful to our team and to NVIDIA, whose partnership helped make this possible. This is what Physical AI from Europe, for the world looks like.
Arne Nordmann, VP Technology & Innovation, NEURA Robotics He was also clear about sequencing.
Although this is a household task, Arne also pointed that this is not only about the task itself. Working together with BSH on this is also to test and start developing machines that robots can easily interact with. NEURA’s first focus stays industry, where demand is high and the operational return is clear. Service robotics follows, and households remain a longer-term goal.
4NE1 Gen 4 walks from the stage to the booth
The keynote ended with a premiere. 4NE1 Gen 4, the new generation of our humanoid robot launched the same day at the opening of the TUM RoboGym in Munich, walked from the keynote stage to the NEURA booth. 4NE1 Gen 4 is designed for series production, is 20 percent lighter than its predecessor, has a slimmer body with a textile outer skin and runs on Neuron AI.
The booth also welcomed NVIDIA founder and CEO Jensen Huang, who stopped by to see the robots at work.
