Machenta

Machenta is a London-based robotics and AI company spun out of Imperial College London. It develops a robot foundation model that, according to the company, learns a new manipulation task from a single demonstration with no further training, and that was trained entirely in simulation without real-world training data. The company calls this capability "Instant Robot Learning".^[1][2] Machenta was co-founded by Edward Johns, Professor of Robot Learning at Imperial and director of its Robot Learning Lab, who is the company's CEO, and his former PhD student Vitalis Vosylius, who is CTO.^[3][4][7] It came out of stealth on 27 October 2026 with a $40 million seed round co-led by Creandum and Union Square Ventures (USV).^[2][3][4]

Business Insider reported at launch that Machenta had 11 employees, remained focused on research and had no near-term plans to commercialize its technology, and its website invites robot owners to register interest in early access to the model.^[3][20]

Overview

Sources for the table: Companies House, the company's team page, Imperial College London and Business Insider.^[9][7][4][3]

History

Research origins at Imperial

Machenta's founders trace the company to research begun at Imperial College London around 2020. Imperial said the work "began six years ago" when Johns and his team at the Robot Learning Lab started investigating whether a robot could learn new physical skills as quickly as a human.^[4] The company's team page says that in 2020 Johns began building a research group dedicated to one-shot robot learning and that the group focused on that one problem for six years.^[7] The company's vision page describes the founders studying one-shot robot learning in a university lab and choosing to pursue "a robot brain that could learn instantly" while much of the field pursued large-scale physical data collection.^[8]

Business Insider described the lab's early experiments: a researcher would physically guide a robot through a task, such as picking up a spatula and scooping an object, and the robot then spent about 30 minutes collecting additional data about the object and its surroundings before it could perform the task on its own. Johns and Vosylius spent the following years trying to remove that extra step.^[3] The lab's 2021 Coarse-to-Fine Imitation Learning paper used a procedure of this kind: after a single demonstration, the robot moved its wrist-mounted camera around the object to collect data for training a pose estimator by self-supervised learning.^[16]

According to Machenta, two 2025 results triggered the company's founding: Instant Policy, which it says showed that in-context learning was possible with pre-training only in simulation, and Learning a Thousand Tasks in a Day, which studied one-shot learning at scale.^[7] Imperial said that after the lab's techniques "began to show strong potential", Johns and Vosylius founded Machenta with colleagues from the lab to scale the work.^[4]

Incorporation and team

Machenta Ltd was incorporated in England on 21 July 2025, with Johns and Vosylius appointed as directors on that date.^[9] Business Insider reported that the founders started the company after seeing early signs that their model could perform new tasks immediately after a demonstration.^[3] In October 2026 the company's team page said it had been building Machenta "for less than a year" and that every author of the five lab papers it lists as its research lineage was now working at the company, alongside researchers from DeepMind, Dyson, Huawei and Stability AI.^[7] The company works from an office and lab in King's Cross, London.^[4][7]

Seed round and launch

On 27 October 2026, Machenta came out of stealth. Johns announced the company and the seed round on X, and Business Insider published the news as an exclusive the same day.^[2][3] Imperial College London and Machenta described the $40 million round as the largest seed round ever raised by a UK robotics startup; this is the company's and the university's characterization.^[4][5] Machenta declined to disclose its valuation.^[3]

Sources for the table: Johns's announcement, Business Insider, Tech Funding News and Imperial College London.^[2][3][6][4]

In statements released with the announcement, Creandum partner Kathi Wilhelm said the team had "spent years breaking through this extremely hard problem, long before there was a company around it",^[4][6] and USV managing partner Rebecca Kaden said the industry was "pouring enormous amounts of capital into collecting real-world data and training robots one task at a time".^[6] USV wrote on X that the approach "allows anyone to teach a robot a task with a single demonstration and have it adapt when things change".^[10] Mary Ryan, Imperial's Vice-Provost (Research and Enterprise), said the investment was a sign that the UK could become a leader in AI-enabled robotics.^[4]

Founders

Johns's Imperial profile says he received Imperial's President's Award for Outstanding Early Career Researcher in 2022, the UK-RAS Early Career Award in 2023, and the Best Conference Paper award at ICRA 2024.^[11]

Instant Robot Learning

Machenta set out its approach in a long launch post, "Introducing Instant Robot Learning", published on 27 October 2026, which the company asks readers to cite as the reference for the model. The description below is the company's own account and has not been independently evaluated.^[1]

The problem Machenta targets

The post contrasts its approach with vision-language-action models (VLAs), which condition on camera images and a language task description and predict robot actions. Machenta argues that because the space of objects, tasks, spatial configurations and environments is effectively infinite, VLA pre-training data is sparse, and a pre-trained model usually has to be post-trained on many real-world demonstrations of each task before deployment, which it says can take days or weeks per task.^[1] Johns summarized the conventional process as "hours/days of data collection and training to teach a robot".^[2]

Vision-demo-action model

Machenta calls its architecture a vision-demo-action (VDA) model. Instead of describing the task in language, the user shows the robot a demonstration. Once trained, the model learns a new task by taking a single demonstration plus a continuous stream of camera images and predicting actions end to end in closed loop, with no change to its weights.^[1] This makes the approach a form of in-context learning applied to imitation learning: the demonstration acts as a prompt, in the way a text prompt conditions a large language model.

According to the post, the training data is generated to capture variation both between tasks and within tasks so that it "induce[s] a meta-learning objective": the weights are optimized to learn how to learn a new task rather than to memorize specific tasks. Because that objective needs very large amounts of data, Machenta generates it in simulation, which it calls its data flywheel. The company says its simulated world is randomized across objects, tasks, configurations, environments, demonstrations and robot embodiments, and that the real-world tasks it shows are not simulations of the same tasks.^[1]

Machenta frames this as an application of Richard Sutton's Bitter Lesson, using compute to learn the model, to learn the task-learning process, and to generate the simulated training world. It says the results published at launch came from "only a modest amount of compute and data scaling so far" and that its next phase is to test whether larger simulated datasets, model capacity and training compute produce more capable instant learners.^[1]

Sim-to-real and language

The usual objection to simulation-trained robot policies is the sim-to-real gap. Machenta argues that the gap mainly hurts policies that learn a specific task in simulation and rely on the simulator's camera parameters or object physics. Its model is instead trained across a wide range of randomized environments to infer what matters from the demonstration it is given, and at test time both the demonstration and the observations come from the real world. In the company's words, "to the model, reality is just one more environment to adapt to".^[1]

The company argues that language is a sparser task description than a demonstration because it says what to do but not how to do it physically. It says its model can be extended to language-conditioned control either by using language to retrieve a matching demonstration (as in Learning a Thousand Tasks in a Day) or by learning a mapping between language and demonstrations (as in Instant Policy), but that its current focus is learning physical behavior.^[1]

The post also names Rhoda AI, Generalist and Skild AI as companies adopting in-context learning for robots, and contrasts their scaling of real-world data collection with Machenta's bet on simulation, reserving real data for the single teaching demonstration.^[1] See Rhoda AI, Generalist AI and Skild AI.

Demonstrations

Machenta's launch post and Johns's announcement thread consist largely of videos. In every video, the company says, the robot had not seen the objects before and learned the task from one demonstration with no further training. The typical video shows a person giving a demonstration, rearranging the objects, stepping back and the robot performing the task autonomously, in a single uncut take.^[1] The videos were selected by the company; Robotics Weekly noted at launch that nobody outside Machenta had yet tested the one-demonstration claim.^[19]

Sources for the table: Machenta and Business Insider.^[1][3]

Physical prompting

Machenta describes several ways of interacting with the model that resemble prompt engineering with a language model:^[1]

• Steerability: with the same set of objects, three different demonstrations produce three different behaviors, which the company presents as evidence that the model conditions on the demonstration rather than recalling a task from its pre-training data.
• Clarifying demonstrations: if a single demonstration is ambiguous (for example, whether an object belongs in "small" containers or "black" containers), further demonstrations can disambiguate it.
• Object generalization: after being taught with a "big white envelope" and a "small brown envelope", the robot picks up a "big brown" and a "medium blue" envelope, which the company calls compositional generalization across shape and color.
• Demo stitching: a long-horizon task can be taught as several sub-demonstrations that are chained together.

Johns calls the resulting workflow "physical vibe coding", a reference to vibe coding.^[2][5] Machenta's example is a "double handover": to pack shoes tightly into a box with parallel-jaw grippers, the team tried several demonstrations until one, in which the shoe is passed between the two arms to flip it, worked reliably. The company says each change to the demonstration immediately changed the robot's behavior.^[1]

Reliability tests

To test reliability, Machenta ran three self-resetting tasks continuously for an hour each over a weekend. It reports 74 of 74 block placements into a bowl, 78 of 78 back-and-forth screwdriver handovers and 97 of 99 attempts at standing a bottle upright, attributing the two failures to a potential collision and kinematic constraints. The company described these tasks as "relatively simple".^[1] Business Insider and Tech Funding News reported that, according to the company, more precise tasks and complex movements remain challenging.^[3][6]

Research background

Machenta's technology builds on the Robot Learning Lab's work on learning manipulation from one or a few demonstrations. The company's team page lists five papers as its lineage.^[7]

Instant Policy

Instant Policy is the closest published precursor to Machenta's model, and Machenta's team page credits Vosylius with building the first generalist in-context learning models for robotics trained entirely in simulation.^[7] The paper was posted to arXiv in November 2024 and accepted as an oral presentation at ICLR 2025.^[12][13] The Robot Learning Lab says it won the Microsoft Best Paper Award at the ICLR 2025 Robot Learning Workshop.^[14]

The method represents demonstrations, the current point-cloud observation and future robot actions together in one heterogeneous graph, and treats in-context imitation learning as a graph generation problem solved with a learned diffusion model that denoises only the action nodes.^[12] Its training data is "pseudo-demonstrations": procedurally generated trajectories in simulation in which a gripper visits randomly sampled waypoints on or near objects drawn from the ShapeNet dataset, grouped so that each set performs the same arbitrary pseudo-task. The authors argue that because the model's weights encode a task-agnostic ability to interpret context rather than specific tasks, such arbitrary data is a "virtually infinite pool" of training data.^[12]

In the paper's experiments, Instant Policy was evaluated on 24 RLBench simulation tasks and 16 real-world tasks with a Sawyer arm and kinesthetic demonstrations. On the real tasks it averaged an 88.75% success rate over 10 rollouts per task, against 34.38% to 40.63% for the baselines tested.^[12] For the real-world experiments the authors co-fine-tuned the model with 5 demonstrations from each of 5 tasks not used in evaluation, to cope with imperfect segmentation and noisy point clouds.^[12] The paper also demonstrated, qualitatively, transfer from human-hand demonstrations and zero-shot execution of tasks specified in language. It listed limitations including reliance on segmented point clouds, no color information, a focus on short-horizon tasks, sensitivity to demonstration quality, no collision avoidance, and insufficient precision for very low-tolerance or contact-rich tasks.^[12]

Machenta's launch post differs from Instant Policy in several described respects: it reports end-to-end predictions from camera images, six robot embodiments and four teaching modes, and it states that the model used no real-world training data at all.^[1] The launch post does not describe the new model's architecture or size.^[1]

Learning a Thousand Tasks in a Day

The Science Robotics paper took a different, retrieval-based route to efficiency. Through 3,450 real-world rollouts, it studied splitting manipulation into an alignment phase and an interaction phase and found that, with fewer than 10 demonstrations per task, this decomposition improved data efficiency by an order of magnitude over single-phase behavioral cloning. The resulting method, MT3, learned everyday tasks from as little as one demonstration each, and the authors used it to teach a robot 1,000 distinct tasks in under 24 hours of demonstrator time.^[18] The paper appeared on the cover of Science Robotics' November 2025 issue, according to the lab.^[14]

Related work elsewhere

Demonstration-conditioned robot policies are an active research area; Generalist GEN-1.5 covers several other in-context approaches. Johns's lab has also worked on retrieval-based one-shot manipulation with vision foundation models: DINOBot (ICRA 2024), by Norman Di Palo and Johns, uses features from Vision Transformers trained with DINO to retrieve the most visually similar object seen in human demonstrations and then align the robot's end-effector with the new object.^[21] The DINOv2 article covers related robotics uses.

Reception

Coverage of the launch focused on the size of the seed round and on Machenta's contrarian choice of simulation over real-world data collection, at a time when a number of companies were paying people to film themselves at work to produce robot training data.^[3][6] Business Insider named Physical Intelligence, Field AI and Skild AI as better-funded rivals, and noted that Skild had unveiled a model in August 2026 that it said could perform unseen tasks after a single video demonstration.^[3] Tech Funding News observed that if simulation alone proves sufficient, robot training may not need as much real-world data, and that investors in both approaches would find out which one works.^[6]

Robotics Weekly, which summarizes robotics posts on X, noted that the launch evidence came from the company's own thread and an investor in the round, and that the thread did not give success rates for the demonstrated tasks; the company's blog published the reliability figures above.^[19][1]

Continue reading

• RRobot foundation modelA robot foundation model is a large-scale machine learning model, typically based on the transformer architecture, that is pre-trained on broad, diverse datasets of robot interactions and then adapted to a…Mentioned twice above
• EEdward JohnsEdward David Johns is a British robotics researcher who is Professor of Robot Learning in the Department of Computing at Imperial College London, where he directs the Robot Learning Lab, and co-founder and CEO…Mentioned 8 times above
• Skild AISkild AI is a robotics artificial intelligence company building a general-purpose foundation model for physical embodimentsMentioned 3 times above
• Rhoda AIRhoda AI (also called Rhoda) is an American robotics company that builds robot foundation models pre-trained mainly on internet video, along with its own robot hardware for factory and warehouse work.Mentioned twice above
• Field AIField AI (also written FieldAI) is an American robotics software company based in Irvine, California, that develops foundation models specifically designed for autonomous robots operating in outdoor and…Also in Embodied AI
• PPhysical IntelligencePhysical Intelligence (also known as Pi or π) is an American artificial intelligence robotics company that builds general-purpose foundation models and learning algorithms designed to control any robot for any…Also in Embodied AI

More to read

• Generalist AI
• Vision-language-action model
• Behavioral Cloning
• Generalist GEN-1.5
• Mimic Robotics
• Mundane Systems

Topics:AI CompaniesEmbodied AIPhysical AIRobotics Companies

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17Machenta, "Introducing Instant Robot Learning", Machenta Blog, 27 October 2026. machenta.com/...instant-robot-learning
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7Edward Johns (@Ed__Johns), announcement of Machenta and its $40M seed round, X, 27 October 2026. x.com/...2115030792720031996
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14Rya Jetha, "An 11-person startup raised $40 million to challenge robotics' data-hungry giants", Business Insider, 27 October 2026. businessinsider.com/...bots-virtual-worlds-2026-10
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10David Silverman, "Imperial AI spinout Machenta raises $40 million to make robots instant learners", Imperial College London News, 27 October 2026. imperial.ac.uk/...-to-make-robots-instant-learners
5. ^1 ^2 ^3John Reynolds, "London robotics startup Machenta raises $40M seed investment", Tech.eu, 27 October 2026. tech.eu/...tup-machenta-raises-40m-seed-investment
6. ^1 ^2 ^3 ^4 ^5 ^6Sofia Chesnokova, "Imperial's Machenta raises $40M from Creandum and USV to train robots in simulation", Tech Funding News, 27 October 2026. techfundingnews.com/...v-simulation-robot-learning
7. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Machenta, "Team". Accessed 30 October 2026. machenta.com/team
8. ^Machenta, "Vision". Accessed 30 October 2026. machenta.com/vision
9. ^1 ^2Companies House, "MACHENTA LTD, company number 16594535". Accessed 30 October 2026. find-and-update.company-information.service.gov.uk/...16594535
10. ^Union Square Ventures (@usv), post on Machenta's seed round, X, 27 October 2026. x.com/...2115075512989086110
11. ^1 ^2Imperial College London, "Professor Edward Johns", profile. Accessed 30 October 2026. profiles.imperial.ac.uk/e.johns
12. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Vitalis Vosylius and Edward Johns, "Instant Policy: In-Context Imitation Learning via Graph Diffusion", arXiv:2411.12633, 2024 (v2 April 2025). arxiv.org/...2411.12633
13. ^1 ^2OpenReview, "Instant Policy: In-Context Imitation Learning via Graph Diffusion", ICLR 2025 Oral. openreview.net/forum
14. ^1 ^2The Robot Learning Lab at Imperial College London, home page and news. Accessed 30 October 2026. robot-learning.uk
15. ^1 ^2Vitalis Vosylius and Edward Johns, "Few-Shot In-Context Imitation Learning via Implicit Graph Alignment", Proceedings of The 7th Conference on Robot Learning, PMLR 229, 2023. proceedings.mlr.press/...vosylius23a
16. ^1 ^2Edward Johns, "Coarse-to-Fine Imitation Learning: Robot Manipulation from a Single Demonstration", ICRA 2021, arXiv:2105.06411. arxiv.org/...2105.06411
17. ^Pietro Vitiello, Kamil Dreczkowski and Edward Johns, "One-Shot Imitation Learning: A Pose Estimation Perspective", CoRL 2023, arXiv:2310.12077. arxiv.org/...2310.12077
18. ^1 ^2Kamil Dreczkowski, Pietro Vitiello, Vitalis Vosylius and Edward Johns, "Learning a Thousand Tasks in a Day", Science Robotics, 12 November 2025, doi:10.1126/scirobotics.adv7594; arXiv:2511.10110. arxiv.org/...2511.10110
19. ^1 ^2 ^3Robotics Weekly, "Machenta exits stealth with $40M and one-demo learning", 27 October 2026. roboticsweekly.news/...-one-demonstration-learning
20. ^Machenta, home page. Accessed 30 October 2026. machenta.com
21. ^Norman Di Palo and Edward Johns, "DINOBot: Robot Manipulation via Retrieval and Alignment with Vision Foundation Models", ICRA 2024, arXiv:2402.13181. arxiv.org/...2402.13181

Show all 21 references