Machenta Raises $40M Seed for Instant Robot Learning

Creandum arrived at Machenta's robotics lab with caps, a roll of paper and a lucky cat. The investor wanted to see what happened when the objects came from outside the demonstration setup, a small but consequential change in who controlled the experiment.

The London robot-learning company announced a $40M Seed round on October 27, 2026, co-led by Creandum and Union Square Ventures. Its founders, CEO Edward Johns and CTO Vitalis Vosylius, are developing a model that Machenta says learns a new physical task from a single demonstration without additional training. The funding supports model scaling and team expansion at a new King's Cross office and lab.

Machenta's commercial proposition begins with the person showing the robot what to do. Making that interaction quick enough to be useful could change how a manufacturer, developer or small team approaches a new automation job, especially when the task changes again before a lengthy training exercise would have paid for itself.

A Seed Round Behind a Different Teaching Process

Machenta comes out of Imperial College London's Robot Learning Lab, where Johns and colleagues spent years investigating one-shot robot learning. The company's team page identifies Johns as co-founder and CEO and Vosylius as co-founder and CTO. It describes a research group that began pursuing this direction in 2020 and brought its work into the startup, giving the financing a longer technical history than the company's brief public life suggests.

The dated Creandum investment announcement confirms that the firm co-led the $40M Seed with Union Square Ventures. Creandum also reports bringing unfamiliar props to the lab and seeing the robot execute a task after a single demonstration, including when objects were moved. That account helps explain the investor's interest in adaptability, while remaining the testimony of a firm with a financial stake in the company.

The round amount is the announced Seed financing; a complete lifetime funding total and valuation have not been established from the primary announcements reviewed. The operational destination of the capital is clearer. Machenta is expanding its team and facilities while increasing the scale of the model, connecting a research claim about learning to the resources needed to develop it further.

How Instant Robot Learning Uses a Demonstration

Machenta calls its approach a vision-demo-action model. In its technical launch, the company describes a system that uses a demonstration to specify the task and a continuing stream of images to guide actions. Its model is pre-trained entirely in simulation, with simulated variation intended to teach the broader process of learning a task from an example.

That distinction matters when interpreting the claim about real-world training data. A real demonstration and real observations still enter the system when a person teaches the robot. According to Machenta, the new task can then be performed without another round of parameter training. The proposed convenience comes from bringing instruction into the interaction itself, making an example useful immediately instead of sending it into a separate task-specific training pipeline.

The launch also describes ways to add demonstrations when an instruction is ambiguous or a longer task needs to be broken into stages. An example can show an object going into a container without making clear whether the important feature is its color, size or position. Machenta's teaching process allows further examples to clarify the intention, leaving the human responsible for expressing a task the robot can actually interpret.

What the Robotics Evidence Establishes

The current launch includes company-produced demonstrations across several robot configurations and teaching modes. Machenta reports that the same pre-trained model is used throughout those videos, including examples of adapting to rearranged objects and reacting to disturbances. These are specific demonstrations of the company's approach, rather than a published guarantee that any physical task can be learned from any example.

The research lineage provides another useful reference point. The earlier Instant Policy project, presented at ICLR 2025, explored in-context imitation learning using simulated pseudo-demonstrations. A separate Learning a Thousand Tasks in a Day project studied retrieval and trajectory decomposition. Those projects help explain the team's focus on learning efficiency, but their methods and results should remain distinct from the current proprietary foundation model.

The commercial interpretation follows from that separation. A team adopting the model would still need to evaluate its own tasks, hardware, teaching conditions and acceptable failure rates. A faster teaching interaction would be valuable because somebody could revise an instruction and see the result sooner; the economics of an actual deployment would also depend on how consistently the robot performs the work after the demonstration ends.

DevCuration's reporting on Maxinsights' physical-AI data business follows the preparation of human recordings for model developers. Its coverage of Formic's factory-robotics expansion examines deployment and ongoing service. Those adjacent businesses show how much work surrounds a robot model, with training data, task teaching and production support creating different supplier responsibilities.

Scaling the Model and the Team in King's Cross

Machenta's next development phase links simulation diversity, model capacity and compute. The company describes broader and more capable instant learners as the objective of scaling those inputs. That is a research direction with an explicit technical mechanism, while the resulting capabilities still need to emerge through experiments and evaluation.

Hiring makes the work behind that ambition more concrete. Machenta lists openings in robot learning, robot simulation, robot control and 3D transformers, alongside community and wildcard roles. Its robot-learning vacancy describes work on representation learning, data quality and diversity, training pipelines, and policies evaluated on real robots. The role is full-time and based in person in London.

Early-access registration is open on Machenta's website, placing interested developers at the beginning of a relationship rather than establishing general commercial availability. For the company, the lab, the training process and the person supplying the demonstration now belong to the same product-development problem. The new King's Cross team will be working through how much of an unfamiliar job a robot can understand while its future user is there to show it.