Machenta builds a foundation model that lets a robot learn a new physical task from one demonstration and then perform it with no further training. The model is pretrained entirely in simulation, with no real-world training data in that set. Someone shows the task once on objects the robot has not seen, and the model predicts actions from that demonstration and a live camera feed. In its launch demos, the same model handled packaging, assembly, and tidying across several robots, including humanoids, from industrial arms to smaller machines on a desk. For a factory task or a job at home, that cuts out the hours to weeks of extra data collection and retraining that still sit between a new task and a working robot. The London company left stealth with a $40 million seed round co-led by Creandum and Union Square Ventures. Edward Johns, professor of robot learning at Imperial College London, is co-founder and CEO, and Vitalis Vosylius, his former PhD student, is co-founder and CTO. The work follows about six years of one-shot and in-context learning research in Johns's lab. Machenta has 11 people, is still focused on research, and has no near-term plan to commercialize, with more precise tasks and complex movements still difficult. The round will expand the team and fund more compute and simulated data at a new office and lab in King's Cross, and the company is opening early access to the model.