Edward Johns

Edward 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 of the robotics startup Machenta.^[1][2][25] His research focuses on robot manipulation, and much of his lab's work since 2020 has addressed how a robot can learn a new task from a single human demonstration, through one-shot imitation learning and, later, in-context imitation learning trained on simulated data.^[1][3][4] Machenta, which builds on that research, came out of stealth on 27 October 2026 with a $40 million seed round co-led by Creandum and Union Square Ventures.^[2][5][6]

Education and career

Johns studied at the University of Cambridge, where he received a BA and an MEng in Electrical and Information Engineering, and then completed a PhD in computer vision at Imperial College London.^[1] His doctoral thesis, "Generative methods for scene association with 2D pairwise constraints" (dated July 2013; Imperial's Spiral repository records the degree as awarded in December 2015), was supervised by Guang-Zhong Yang. It addressed recognizing which known place a query image shows, for robot navigation tasks such as loop closure, global localization and topological navigation, and its experiments used a dataset collected by walking a 7 km route through a park and urban district over eight months.^[7] His papers from this period, on appearance-based place recognition and long-term localization, were co-authored with Yang and appeared at venues including ICCV 2011, ICRA 2013 and the International Journal of Computer Vision.^[8]

After the PhD he spent a year as a postdoctoral researcher at University College London.^[1] A CVPR 2015 paper from that period, "Becoming the Expert: Interactive Multi-Class Machine Teaching", lists him with UCL's Oisin Mac Aodha and Gabriel Brostow.^[8] He then returned to Imperial as a founding member of the Dyson Robotics Lab, led by Andrew Davison, where he headed the robot manipulation team.^[1] Imperial announced the lab in February 2014 as a £5 million collaboration with Dyson.^[9][10]

In 2018 Johns received a Royal Academy of Engineering Research Fellowship, which his Imperial profile attributes to his work on sim-to-real transfer and imitation learning. He then joined Imperial's faculty and founded the Robot Learning Lab.^[1] He was also Head of Robot Learning at Dyson from June 2021 to June 2022.^[1]

Sources for the table: Imperial College London profile, Spiral thesis record, Dyson Robotics Lab, Machenta and Companies House.^{[1][7][9][11][25]} The date of his promotion from Reader to Professor is not given on his Imperial profile. Imperial's October 2026 announcement of Machenta calls him Professor of Robot Learning.^[2]

Research

The Robot Learning Lab describes its research as lying at the intersection of robotics, computer vision and machine learning, with a primary focus on robot manipulation. It names imitation learning, reinforcement learning and vision-language models as its main lines of investigation, with applications in domestic, manufacturing and warehouse robots.^[3] Johns's Imperial profile says he has published more than 80 peer-reviewed papers with more than 8,000 citations.^[1]

Early work: grasping and sim-to-real transfer

At the Dyson Robotics Lab, Johns worked on learning robot skills from simulated data. "Deep Learning a Grasp Function for Grasping under Gripper Pose Uncertainty" (IROS 2016, with Stefan Leutenegger and Andrew Davison) trained a CNN on grasp attempts collected in simulation, and "Pairwise Decomposition of Image Sequences for Active Multi-View Recognition" (CVPR 2016 oral) addressed multi-view object recognition.^[4] In "Transferring End-to-End Visuomotor Control from Simulation to Real World for a Multi-Stage Task" (CoRL 2017), Stephen James, Davison and Johns trained a controller to grasp a cube and drop it into a basket using behavioral cloning in simulation, then transferred it to a real robot with domain randomization.^[4][12] Later lab papers benchmarked sim-to-real transfer for dynamics (IROS 2020) and domain randomization for visual pose estimation (ICRA 2021).^[4]

One-shot imitation learning

From 2021 the lab published a series of methods that learn a manipulation task from a single demonstration. Johns's single-author ICRA 2021 paper, "Coarse-to-Fine Imitation Learning: Robot Manipulation from a Single Demonstration", splits a task into a coarse approach trajectory and a fine interaction trajectory. A state estimator for the pose at which interaction begins is trained by self-supervised learning, with the robot moving its wrist camera around the object; at test time the robot moves to that pose and replays the demonstration's end-effector velocities. The paper reports experiments on 8 everyday tasks.^[13] Machenta's team page calls a 2021 self-supervised learning paper the one "that started it all".^[11]

DOME ("Demonstrate Once, Imitate Immediately", IROS 2022) needs no data collection after the demonstration. It uses an object segmentation network and a visual servoing network, trained in advance in simulation, so that a task can be performed straight after one demonstration with no further data collection or training; the authors report near 100% success on 7 real-world tasks.^[4][14] Later work studied the problem from a pose estimation perspective (CoRL 2023), used DINO vision features to retrieve and align demonstrations for novel objects (DINOBot, ICRA 2024; see DINOv2), and extended one-shot learning to dual-arm tasks (ICRA 2025).^[4][15]

In a CoRL 2021 blue sky paper, "Back to Reality for Imitation Learning", Johns argued that evaluation of real-world robot learning should focus on time efficiency, which better reflects the cost to humans, rather than the data efficiency metrics inherited from machine learning.^[4][16]

In-context imitation learning

The lab's later work treats demonstrations as a prompt for a pre-trained model, a robotics counterpart to in-context learning in large language models. "Few-Shot In-Context Imitation Learning via Implicit Graph Alignment" (CoRL 2023, with Vitalis Vosylius) generated large numbers of virtual demonstrations by deforming random objects in simulation.^[4] "Keypoint Action Tokens Enable In-Context Imitation Learning in Robotics" (RSS 2024, with Norman Di Palo) represented observations and actions as 3D keypoints so that demonstrations could be fed to a large language model, using the model's pattern recognition for low-level action prediction rather than high-level reasoning in natural language.^[4][17]

"Instant Policy: In-Context Imitation Learning via Graph Diffusion" (Vosylius and Johns) learns new tasks from one or two demonstrations without further training. It represents demonstrations, observations and actions as a graph, generates actions with a learned diffusion process, and is trained on "pseudo-demonstrations": arbitrary trajectories generated in simulation.^[18] It was presented as an oral at ICLR 2025 and, according to the lab, won the Best Paper Award at the ICLR 2025 Robot Learning Workshop.^[3][19]

"Learning a Thousand Tasks in a Day" (Kamil Dreczkowski, Pietro Vitiello, Vosylius and Johns), published in Science Robotics in 2025, compared ways of decomposing manipulation into alignment and interaction phases across 3,450 real-world rollouts. The resulting method, Multi-Task Trajectory Transfer (MT3), learns tasks from as little as one demonstration each; the paper reports teaching a robot 1,000 distinct everyday tasks in under 24 hours of human demonstrator time.^[20] The lab says the paper appeared on the journal's front cover.^[3] Machenta names Instant Policy and Learning a Thousand Tasks in a Day as the two 2025 results that triggered the company's founding.^[11]

Language and generative models

The lab has also worked with web-scale generative models. DALL-E-Bot (RA-L 2023) inferred a text description of a scene, prompted DALL-E to generate a goal image, and rearranged objects to match it; the lab calls it the first work to study web-based diffusion models for robotics.^[4][21] Dream2Real (ICRA 2024) built a NeRF of a scene, rendered candidate rearrangements and scored them against a language command with CLIP.^[4] Johns is a co-author of Prismer (TMLR 2024), a vision-language model with multi-task experts, with researchers including Linxi Fan and Anima Anandkumar.^[4]

Johns is one of the many co-authors of "Open X-Embodiment: Robotic Learning Datasets and RT-X Models", the cross-institution Open X-Embodiment dataset paper, which won the Best Conference Paper Award at ICRA 2024.^[22]

Selected publications

Sources for the table: the Robot Learning Lab publication list and the arXiv records of each paper.^{[4][12][13][14][15][16][17][18][20][21][23][26]}

Machenta

Johns co-founded Machenta with Vosylius, his former PhD student, who is the company's CTO.^[6][11] Companies House records Johns's appointment as a director of Machenta Ltd on 21 July 2025.^[25] Imperial says the two founded it with colleagues from the Robot Learning Lab to scale the lab's work.^[2][6] The company says every author of the five lab papers it lists as its research lineage (from the 2021 self-supervised learning paper to Instant Policy and Learning a Thousand Tasks in a Day) now works at Machenta.^[11] Business Insider reported at launch that the company had 11 employees and was focused on research, with no near-term plans to commercialize its technology.^[6]

Johns announced the company on X on 27 October 2026, together with the $40 million seed round co-led by Creandum and Union Square Ventures.^[5] Machenta describes its model as a "vision-demo-action" model that learns a new task from one demonstration with no further training and was trained entirely in simulation; it calls the capability "Instant Robot Learning", and Johns described teaching robots with it as "physical vibe coding". These are the company's claims, shown through its own videos.^[5][24] Imperial described the round as the largest seed raise by a UK robotics startup.^[2] Speaking to Imperial at the launch, Johns said: "Intelligence is not so much about what you can already do, it's about how quickly you can learn to do something new."^[2]

Awards and recognition

His Imperial profile also says he has served on the advisory boards of several robotics and AI startups.^[1]

Continue reading

• IImitation LearningImitation learning is a family of methods for learning sequential behavior from demonstrations.Mentioned 20 times above
• MMachentaMachenta is a London-based robotics and AI company spun out of Imperial College London.Mentioned 12 times above
• DINOv2DINOv2 is a family of self-supervised Vision Transformer models released by Meta AI Research in April 2023 that produces general-purpose visual features transferring to many downstream tasks without…Links back to this article
• RRobot manipulationRobot manipulation is the ability of a robotic system to physically interact with objects in its environment through grasping, pushing, pulling, inserting, placing, and other contact-based actions.Mentioned 7 times above
• SSim-to-real transferSim-to-real transfer (also written as sim2real) is the process of training a policy, controller, or perception model inside a physics simulator and deploying it on a physical robot in the real world without…Mentioned 3 times above
• OOpen X-EmbodimentOpen X-Embodiment (OXE) is a large-scale collaborative robotics research initiative, led by Google DeepMind and announced in October 2023, that produced the largest open-source real robot dataset and a family…Mentioned 3 times above

More to read

• Jim Fan
• Humanoid Robot
• Vision-language-action model
• Dexterous hand
• Teleoperation
• Physical AI

Topics:Embodied AIPeopleRobotics

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13Imperial College London, "Professor Edward Johns", staff profile (bio, academic and non-academic positions, degrees). Accessed 30 October 2026. profiles.imperial.ac.uk/e.johns
2. ^1 ^2 ^3 ^4 ^5 ^6David 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
3. ^1 ^2 ^3 ^4 ^5The Robot Learning Lab at Imperial College London, home page and latest news. Accessed 30 October 2026. robot-learning.uk
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13The Robot Learning Lab at Imperial College London, "Our Research" (publication list with mini abstracts). Accessed 30 October 2026. robot-learning.uk/research
5. ^1 ^2 ^3Edward Johns (@Ed__Johns), announcement of Machenta and its $40M seed round, X, 27 October 2026. x.com/...2115030792720031996
6. ^1 ^2 ^3 ^4Rya 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
7. ^1 ^2Edward David Johns, "Generative methods for scene association with 2D pairwise constraints", PhD thesis, Department of Computing, Imperial College London, 2013 (advisor Guang-Zhong Yang), Spiral repository, doi:10.25560/33210. hdl.handle.net/...33210
8. ^1 ^2OpenAlex, author record for Edward Johns (ORCID 0000-0002-8914-8786), works list. Accessed 30 October 2026. openalex.org/A5080385974
9. ^1 ^2Imperial College London, Dyson Robotics Lab, "About us". Accessed 30 October 2026. imperial.ac.uk/...about-us
10. ^Imperial College London, "Dyson and Imperial to develop next generation robots at new centre", 10 February 2014. imperial.ac.uk/...l-develop-next-generation-robots
11. ^1 ^2 ^3 ^4 ^5Machenta, "Team". Accessed 30 October 2026. machenta.com/team
12. ^1 ^2Stephen James, Andrew J. Davison and Edward Johns, "Transferring End-to-End Visuomotor Control from Simulation to Real World for a Multi-Stage Task", CoRL 2017, arXiv:1707.02267. arxiv.org/...1707.02267
13. ^1 ^2Edward Johns, "Coarse-to-Fine Imitation Learning: Robot Manipulation from a Single Demonstration", ICRA 2021, arXiv:2105.06411. arxiv.org/...2105.06411
14. ^1 ^2Eugene Valassakis, Georgios Papagiannis, Norman Di Palo and Edward Johns, "Demonstrate Once, Imitate Immediately (DOME): Learning Visual Servoing for One-Shot Imitation Learning", IROS 2022, arXiv:2204.02863. arxiv.org/...2204.02863
15. ^1 ^2Norman 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
16. ^1 ^2Edward Johns, "Back to Reality for Imitation Learning", CoRL 2021 blue sky oral track, arXiv:2111.12867. arxiv.org/...2111.12867
17. ^1 ^2Norman Di Palo and Edward Johns, "Keypoint Action Tokens Enable In-Context Imitation Learning in Robotics", RSS 2024, arXiv:2403.19578. arxiv.org/...2403.19578
18. ^1 ^2Vitalis Vosylius and Edward Johns, "Instant Policy: In-Context Imitation Learning via Graph Diffusion", arXiv:2411.12633. arxiv.org/...2411.12633
19. ^The Robot Learning Lab at Imperial College London, "Instant Policy" project page. Accessed 30 October 2026. robot-learning.uk/instant-policy
20. ^1 ^2Kamil Dreczkowski, Pietro Vitiello, Vitalis Vosylius and Edward Johns, "Learning a Thousand Tasks in a Day", Science Robotics 10, eadv7594 (2025), doi:10.1126/scirobotics.adv7594; arXiv:2511.10110. arxiv.org/...2511.10110
21. ^1 ^2Ivan Kapelyukh, Vitalis Vosylius and Edward Johns, "DALL-E-Bot: Introducing Web-Scale Diffusion Models to Robotics", IEEE RA-L, arXiv:2210.02438. arxiv.org/...2210.02438
22. ^1 ^2ICRA 2024, "Awards and Finalists" (IEEE ICRA Best Conference Paper Award winners). Accessed 30 October 2026. 2024.ieee-icra.org/awards-and-finalists
23. ^Yilong Wang, Cheng Qian and Edward Johns, "Instant-Fold: In-Context Imitation Learning for Deformable Object Manipulation", accepted at CoRL 2026, arXiv:2606.04269. arxiv.org/...2606.04269
24. ^Machenta, "Introducing Instant Robot Learning", Machenta Blog, 27 October 2026. machenta.com/...instant-robot-learning
25. ^1 ^2 ^3Companies House, "MACHENTA LTD (16594535): officers". Accessed 30 October 2026. find-and-update.company-information.service.gov.uk/...officers
26. ^Ivan Kapelyukh, Xiaohan Zhang, Stephen James, Laura Herlant and Edward Johns, "ZeroBot: Learning from Scratch in Minutes with Generative Real2Sim", IEEE RA-L 2026, arXiv:2609.34010. arxiv.org/...2609.34010

Show all 26 references