Professor of Robot Learning at Imperial College London.
London, UK
Joined January 2019
- Tweets439
- Following147
- Followers4K
- Likes1K
Pinned Tweet
I'm very excited to finally announce one of the most ambitious projects we've worked on — which makes the front cover of Science Robotics today:
☀️ Learning a Thousand Tasks in a Day ⭐️
Everyday tasks — like those below — can now be learned from a single demonstration each...
This is Ivan’s final paper of his outstanding PhD, a thesis which studies numerous creative ways to bring generative AI into robotics. More from Ivan below!
Simify: use generative real2sim to build a simulator, then perform spatial reasoning using evolutionary search. Presented by @YafeiHuCMU at IROS!
arxiv.org/abs/2609.33982
Ivan’s ZeroBot paper is at IROS this week — reconstructing real objects and then learning how to interact with those objects in simulation — take a look below!
ZeroBot: learning from scratch with generative real2sim and fast parallel RL:
arxiv.org/abs/2609.34010
Joint Imperial & RAI internship work, published in RA-L and presented at IROS. Thank you to my wonderful collaborators!
@XiaohanZhang220 @stepjamUK Laura Herlant @Ed__Johns
Great blog post by Chris on in-context and one-shot learning.
This is an area I’ve been exploring for many years, and it’s exciting to see the momentum building now!
In-Context Learning Results Hint at a “GPT Moment” for Robotics
For general-purpose robots to be useful and economical, you need to be able to teach them new skills on the fly. And now we can see how this will work, with in-context long horizon video demonstrations shown by companies like Generalist, Skild, Rhoda, Rhoda, and RobbyAnt.
The idea is simple: prompt with a demonstration, mixed video and proprioception information. Language isn't enough -- it doesn't specify the problem well enough. And it's starting to really work.
more in blog post
Looking forward to this debate!
🎙️ Save the date: a special debate at CoRL 2026, with seven leading voices in robot learning: corl.org/program/special-pan…
🗓️ Nov 11 | 4–5pm | Austin, TX
Topic reveal soon 👀
#CoRL2026 #RobotLearning
Edward Johns retweeted
A robot is supposed to be a tool which one can use to perform a wide variety of tasks. We usually focus on the dream of generalization, but it's interesting instead to focus on that fundamental goal: how can I let someone teach a robot anything I want, quickly?
In "learning a thousand tasks in a day," the authors do just that, using a combination of decomposition and retrieval to get strong generalization with very little data and without training a huge VLA.
Learn more ->
For general-purpose robots to be useful, we must be able to teach them new tasks quickly and repurpose them to new roles. But most techniques for teaching robots — even from human demonstrations — take a great deal of new data. But it turns out that it is possible, in part through decomposing tasks into component subtasks, that it’s possible to learn a robot skills quickly.
@Kamil__Dre @pitvit_ and @Ed__Johns join us to talk about their extensive study on how to teach robots new skills efficiently, while also generalizing to novel object instances, using a combination of task decomposition and retrieval, work that was published in Science Robotics.
Watch episode #94 of RoboPapers now, with @micoolcho and @DJiafei, to learn more!
Edward Johns retweeted
Really enjoyed this podcast with @Ed__Johns and team on training more effectively and efficiently. Maybe the next big thing should be who can train with the least data, not the most.
For general-purpose robots to be useful, we must be able to teach them new tasks quickly and repurpose them to new roles. But most techniques for teaching robots — even from human demonstrations — take a great deal of new data. But it turns out that it is possible, in part through decomposing tasks into component subtasks, that it’s possible to learn a robot skills quickly.
@Kamil__Dre @pitvit_ and @Ed__Johns join us to talk about their extensive study on how to teach robots new skills efficiently, while also generalizing to novel object instances, using a combination of task decomposition and retrieval, work that was published in Science Robotics.
Watch episode #94 of RoboPapers now, with @micoolcho and @DJiafei, to learn more!
Thanks for the fun chat!
Full episode dropping soon!
Geeking out with @Kamil__Dre @pitvit_ @Ed__Johns on Learning a Thousand Tasks in a Day robot-learning.uk/learning-1…
Co-hosted by @micoolcho @DJiafei
Congratulations to George Papagiannis (@geopgs) on passing your PhD viva today! 🍾
His thesis is so colourful that I made him a tea towel out of all the figures...
Check out George's wonderful papers here:
robot-learning.uk/adapting-s…
robot-learning.uk/miles
robot-learning.uk/r-plus-x
Great article here by Jack — exploring the factors at play when answering: “At what point does a robot become economically worthwhile?”
London — let’s go!
Anthropic has announced that it is massively expanding its London presence.
It’s just secured a new office for 800 people - a huge jump from its 200 current employees.
OpenAI announced its first permanent office in London this week and now @AnthropicAI is doubling down.
Meta, OpenAI, DeepMind, wayve and so many others have huge offices in London.
It’s becoming the leading AI hub outside of the US.
LETS GO
The UK Robot Manipulation Workshop has been buzzing today in Edinburgh!
I’m very happy that we’ve helped to build up such a thriving community through these workshops, since we started back in 2016 👌
Thanks to Barbara Webb and Mohan Sridharan for hosting us this year!
As we're building up to 2026, I've taken a moment to look back on all the work from our lab at Imperial in 2025.
For a quick summary, I recorded a short video here: youtu.be/_DEeqKNwTNw.
Thanks to all my team here for a fun and productive year — and here's to a great 2026!
It's happening — we're holding the 7th UK Robot Manipulation Workshop in Edinburgh on January 6th and 7th 🇬🇧🤖🦾
Registration is still open...
And the poster submission deadline is... today!
robot-manipulation.uk
Below is a brief history of how we've grown to 300 attendees:
We began in 2016, with Jeremy Wyatt gathering together ~50 people from the UK robot manipulation community. It was such a wonderful initiative by Jeremy that I then decided to continue the story, by bringing the workshop to Imperial in 2017. In 2019, Mehmet Dogar took the workshop to Leeds, and after a break during the COVID years, @nathanlepora helped to turn this into an annual tradition by hosting the workshop in Bristol in 2023. In 2024, @IngmarPosner and @perla_maiolino invited us all to Oxford for the fifth workshop, and earlier this year @ShanLuoRobotics and @mhoward3210 assembled a mighty ~300 people together for the 2025 edition!
The two-day workshop is an excellent example of grassroots community building across both academia and industry. We've grown steadily from 50 people to 300 people over 9 years, and we're organised by teams of volunteers at universities, with generous support from industry sponsorship.
And this is not like a typical conference where you might only turn up if you have a paper... It's a place to present whatever your latest work is (even if already published elsewhere) in a relaxed setting, as well as a chance to meet in person with other collaborators you might otherwise only see on video calls, and an opportunity to just spend a couple of days hanging out with the rest of our charming community!
We hope to see you there in January, and thank you to Barbara Webb and Mohan Sridharan for bringing this next one to Edinburgh! 🙂
@UK_Robot_Manip
Congratulations, Stephen!
Huge news! We've secured funding to scale @Neuracore_AI and open up the platform to academia. If you work in academia come and sign up for free!
Do you want to see if your robot can learn a thousand tasks in a day ... ?
Following our Science Robotics paper last week, we now give you:
>> Our Code and Model Weights 🧑💻
>> An Explainer Video for the Paper 📽️
For both: robot-learning.uk/learning-1…
Let us know how you get on!
Thank you @Kamil__Dre and @pitvit_ for putting these together 🙂
Here's some behind-the-scenes footage from our Science Robotics paper:
Kamil (@Kamil__Dre) and Pietro (@pitvit_) really did teach our robot 1000 tasks, and — just for fun — we filmed everything!
Full video at: robot-learning.uk/learning-1…
So... we now have 2 challenges for you: 👇
1. (Easy) Can you find when the gripper breaks and they have to fix it ... ?
2. (Hard) Can you find when I come into the lab and provide one of the demos myself ... ? It definitely happened, but honestly, even I haven't been able to find it yet because the video is so long (representing ~17 hours). You'll have to look at the full video on the website — let me know if you can find it!
Edward Johns retweeted
Advent of Visual GenAI has reduced the (hungry & starved) robotics data needs from to only distribution. Which means robots need to explore only the dist, & can be generated.
@Ed__Johns group over the years are shrinking it from dist to a single trajectory!! That means you can train robots on the job!
Edward Johns retweeted
For robots to be useful in homes and small businesses, they'll be expected to do a wide range of tasks. That means using just a couple demonstrations so that the robot user can teach the robot, instead of relying on data farms. Really cool work on how this can be feasible
Our full paper dives deep into the question of:
"Decomposed policies vs monolithic policies — when should you use which?"
For example, how does each scale with the number of demonstrations per task?
This was jointly led by Kamil Dreczkowski (@kamildre) and Pietro Vitiello (@pitvit_), supported by Vitalis Vosylius (@vitalisvos19).
Well done guys, what a fun project!
For many more results and insights, please visit: robot-learning.uk/learning-1….