Roboter lernen von Menschen.Wir bauen die Plattform dafür.

Vision-language-action models do not learn a task from a description. They learn from episodes: a human moves the robot arm through the task while cameras and joint encoders record every frame. A single policy for one gripper task needs dozens of episodes, and every episode has to be recorded on a real arm by a real person. Labs solve this with students and patience, which does not scale. That is the bottleneck: the models are available as open weights today, an arm costs a few hundred euros, but demonstration data is still produced by hand, task by task.

AY-Robots puts the four pieces together. Arm owners connect an SO-100, SO-101, Koch or WidowX via the desktop client and release time slots. Operators take over the arm via WebRTC and record episodes in LeRobot v2.1 format with up to five cameras at 30 fps, at country rates of 8 to 22 euros per hour with a 20 percent platform share. Finished datasets sell on the marketplace at a 10 percent commission. The platform then trains ACT, SmolVLA, Pi0.5 and GR00T on rented GPUs and serves the checkpoint as a cloud endpoint driving a real arm, billed by the minute.

All of that runs today: a real SO-100 is controllable at /live, the client records and uploads, the training pipeline covers five policy architectures, and the MCP server lets Claude or ChatGPT operate the arm directly. What is missing is capacity. Round two therefore goes into our own robot cells, reserved GPU capacity and more arms and operators, so that a company can order a dataset for its task and get a working policy back without building a lab first. The goal is the data and compute layer robot learning needs, as speech models needed transcribed audio.