Machenta raises $40M seed and unveils one-demonstration robot learning model

Date: 2026-10-27
Type: ARTICLE
Topics: Funding, Imitation Learning, Robot Foundation Models, Sim To Real Transfer
Deck: The London startup pairs $40 million of seed funding with a simulation-trained model demonstrated across six robot embodiments.
URL: https://korthosrobotics.com/research/machenta-40m-seed-instant-robot-learning

Image: Machenta
Machenta has raised a $40 million seed round led by Creandum and Union Square Ventures and introduced a robot foundation model designed to learn unfamiliar tasks from a single demonstration. The London company was founded by Imperial College London professor Edward Johns and his former PhD student Vitalis Vosylius. The financing will support model development and expansion of its team and laboratory in King's Cross, according to Imperial's funding announcement.
The company's Instant Robot Learning release describes a Vision–Demo–Action, or VDA, model that receives a demonstration of the desired behaviour alongside camera observations and generates robot actions in a closed loop. Machenta says pretraining uses entirely simulated data. At deployment, the demonstration and subsequent camera images come from the physical environment, allowing the pretrained model to respond to a new task without task-specific fine-tuning or weight updates.
The release shows the same model running on six robot embodiments with different cameras and hardware. Teaching methods include leader–follower control, VR controllers, physical guidance and demonstrations using human hands. One video shows three robots with different grippers and teaching methods performing an assembly workflow after roughly two minutes of instruction. Other demonstrations show responses to rearranged objects and disturbances, tool handling and preliminary control of humanoid hands.
In company-run, one-hour tests on relatively simple self-resetting tasks, the robots completed 74 block placements from 74 attempts, 78 screwdriver handovers from 78 attempts and 97 bottle-uprighting attempts from 99. Machenta attributes the two bottle-task failures to a potential collision and kinematic constraints. The tests give specific conditions and failure examples alongside the other task demonstrations.
The approach builds on six years of research at Imperial, including Implicit Graph Alignment, presented at CoRL 2023, and Instant Policy, presented at ICLR 2025. Machenta's next development phase will increase the diversity of simulated experience, training compute and model capacity to extend the range and complexity of tasks that can be taught through demonstrations.