$40m · Seed · AI · London, UK
Machenta co-founders Edward Johns and Vitalis Vosylius. Image: Oscar May, via Creandum.
London robotics startup Machenta has raised $40m in Seed funding, co-led by Creandum and Union Square Ventures. The company emerged from stealth with a robot foundation model designed to learn a new physical task from one demonstration, and says the capital will support technical scaling and team growth.
Moving robot training from task data to compute
Machenta is trying to remove a costly step in deploying robots: collecting demonstrations and retraining a model for each new job. Its model takes a demonstration, rather than a language instruction, as the task prompt and predicts actions from that demonstration plus a live stream of images. The company says the same pre-trained model has learned packaging, assembly and tidying tasks across six robot types without task-specific fine-tuning.
The technical wager is that simulation can supply enough varied training data for the model to learn how to learn, while the only real-world data needed at deployment is the operator's demonstration. Machenta says its model was trained entirely in simulation, drawing on six years of Imperial College research. In the company's framing, that shifts the scaling constraint away from repeated real-world data collection and towards compute and the quality of simulated environments.
Commercially, the proposition matters because task set-up can determine whether flexible automation pays. If a factory can teach a new task in minutes instead of commissioning another data-collection and training cycle, the same robot can be useful across higher-mix work where conventional automation is harder to justify. The immediate feedback loop also lets an operator revise the demonstration when a robot's first interpretation is wrong, rather than waiting for another training run.
Reliability is the next proof point
The public evidence is still centred on Machenta's own demonstrations and investor diligence, rather than named customer deployments. The company reports that its model completed 74 of 74 block placements, 78 of 78 screwdriver handovers and 97 of 99 bottle-standing attempts in simple, self-resetting tests. Creandum says its team also brought unfamiliar props to the lab and watched the robot adapt after one demonstration.
Those results support the learning mechanism, but industrial value will depend on reliability across longer tasks, safety constraints and changing production environments. The $40m round gives Machenta room to scale compute, simulation data and its roughly dozen-person team from a new King’s Cross base. The milestone that follows the funding is not another isolated demo: it is showing that instant teaching can shorten real deployment without moving the integration burden elsewhere.
Explore this theme
Industrial robotics is a deployment problem, not just an AI challenge — EUVC Summit conversation · 15 October 2026. Sunrise Robotics' production experience explains why faster robot programming still has to meet reliability, installation and operating-economics tests.
