Antoine Cully
@CULLYAntoine
Followers
2K
Following
1K
Media
63
Statuses
637
Professor in Machine Learning and Robotics and Director of the Adaptive and Intelligent Robotics Lab (AIRL) at Imperial College London.
London, England
Joined December 2011
The new ICARL seminar series season has started with as always fantastic speakers! Join us for the next seminar on the 26th of November!
0
0
5
Can a game really talk back? We built The Oversight Bureau to find out : a narrative experiment powered by our on-device AIs. The public demo is live today. Will you challenge The Bureau? https://t.co/aVLQ7mfyoa
store.steampowered.com
The Oversight Bureau is a voice-driven narrative adventure game. You are candidate 404, trapped inside C.U.T.E. - a rehabilitation program designed to make you a more โappropriateโ member of society....
1
2
9
Of first set of guest speakers are set ๐งโ๐ฌ - Alyssa Adams - Alexander Mordvintsev @zzznah - Angel Goรฑi-Moreno - Eric Medvet @EricMedvetTs - Kyrre Glette @kyrre2000 - Stefano Nichele @stenichele - Susan Stepney
0
7
20
A new season of the ICARL seminars begins๐ฅณ๐ฅณ This year we are kicking off with @max_nlp from @GoogleDeepMind! Max will be sharing his insights through various works with LLMs and RL. ๐๏ธOct 17th, 5 pm ๐Huxley Building, Imperial College London, room 311 @ICComputing This is
1
3
7
Excited to present our latest paper "Time to Play"! It's the kind of paper I love: a simple idea that shows some powerful results. What if all the energy of children, which progressively decreases as they grow up, is an essential part of the curriculum? Turns out that's the case!
What if your robot was just missing a childhood to walk faster? Introducing SMOL, for Scaling Mechanical Output over Lifetime, tomorrow (Oct 9) at #ALIFE25! with @CULLYAntoine, @hannah_janmo, and D. Labonte from @EvoBiomech
0
1
5
If you are at #CoRL this week, check out our latest work on unsupervised open-ended discovery of skills directly on real robot, with no priors (even no simulators!). This is achieved by combining world models and Unsupervised Quality Diversity.
We're presenting URSA today! Feel free to visit @lisa_coiffard and I at Poster 75 โก๏ธ #CoRL2025
0
0
8
We're presenting URSA today! Feel free to visit @lisa_coiffard and I at Poster 75 โก๏ธ #CoRL2025
๐คWhat if robots could discover diverse abilitiesโall without simulation nor extensive human tuning? With @lisa_coiffard, Oscar Pang, @maxencefaldor and @CULLYAntoine, we introduce URSA: an efficient skill discovery algorithm applicable directly on real hardware. #CoRL2025 ๐งต
0
2
8
As a newly appointed ๐๐๐๐ถ๐๐๐ฎ๐ป๐ ๐ฃ๐ฟ๐ผ๐ณ๐ฒ๐๐๐ผ๐ฟ at @imperialcollege, I'm thrilled to announce the ๐ฆ๐ฎ๐ณ๐ฒ ๐ช๐ต๐ผ๐น๐ฒ-๐ฏ๐ผ๐ฑ๐ ๐๐ป๐๐ฒ๐น๐น๐ถ๐ด๐ฒ๐ป๐ ๐ฅ๐ผ๐ฏ๐ผ๐๐ถ๐ฐ๐ ๐๐ฎ๐ฏ (๐ฆ๐ช๐๐ฅ๐) at ๐๐บ๐ฝ๐ฒ๐ฟ๐ถ๐ฎ๐น ๐๐ผ๐น๐น๐ฒ๐ด๐ฒ ๐๐ผ๐ป๐ฑ๐ผ๐ป. ๐ฆ๐ฎ๐ณ๐ฒ ๐ช๐ต๐ผ๐น๐ฒ-๐ฏ๐ผ๐ฑ๐
14
14
177
Very impressive work!!
In-context learning for Robotics is here. Skild Brain adapts to new bodies and survives extreme shifts โ without ever being trained on any of these robots.
0
0
4
Open-endedness, quality-diversity, and AI-Generating algorithms++ Exciting to see the scientific progress this will catalyze๐๐ก๐ฅ๐๐งช๐งฌ
Proud to release ShinkaEvolve, our open-source framework that evolves programs for scientific discovery with very good sample-efficiency! ๐ Paper: https://t.co/05rjMiSxOL Blog: https://t.co/yjuvV0xpif Project:
2
2
43
"Always reasoning" (ReAct) isn't optimal for LLM agents! ๐ง Our new paper identifies a "Goldilocks" effect: planning too frequently or not enough degrades performance. We show how to train agents to learn to dynamically allocate test-time compute when needed for best results. ๐
Almost all agentic pipelines prompt LLMs to explicitly plan before every action (ReAct), but turns out this isn't optimal for Multi-Step RL ๐ค Why? In our new work we highlight a crucial issue with ReAct and show that we should make and follow plans instead๐งต
2
20
93
Almost exactly 10 years after joining @imperialcollege as a Postdoc, I am honoured to announce that I am now Professor in Machine Learning and Robotics! ๐จโ๐ ๐ค My fantastic team found the best gift to celebrate this special occasion!
32
7
241
What happens when you combine world models and open-ended skill discovery directly on physical robots?๐ค You get robots that can discover their abilities in a few hours, no simulator needed! Here our robot learns to walk in just 5h! Check out URSA in our latest #CORL2025 paper!
๐คWhat if robots could discover diverse abilitiesโall without simulation nor extensive human tuning? With @lisa_coiffard, Oscar Pang, @maxencefaldor and @CULLYAntoine, we introduce URSA: an efficient skill discovery algorithm applicable directly on real hardware. #CoRL2025 ๐งต
0
2
35
Our first game at @iconicgamesio The Oversight Bureau was teased in NVIDIA's Gamescom update! https://t.co/IfPGAzWhJd Be mindful of what you say around NPCsโthey'll be listening... Their responses and your story will adapt accordingly... All running on-device๐ Come play our
3
4
27
We have invited experts and researchers to test Genie 3 and provide feedback as part of our research preview. Thank you @CULLYAntoine for your time and thoughts!
I have been fortunate and honoured to be invited by @GoogleDeepMind to test an early research prototype of Genie 3 and share my thoughts! The potential of this technology is truly incredible and opens up so many new possibilities! More thoughts below:
1
1
10
Thanks @CULLYAntoine for coming to try Genie 3 and for the super insightful thoughts on where itโs at for creative and robotics use cases. Things are about to get really exciting ๐
I have been fortunate and honoured to be invited by @GoogleDeepMind to test an early research prototype of Genie 3 and share my thoughts! The potential of this technology is truly incredible and opens up so many new possibilities! More thoughts below:
0
3
12
It is also a new creative medium for rapid ideation and quick prototyping, as described by @togelius in his blog. I am simply curious to see what the next steps and use cases for Genie will be! https://t.co/ZHlL5BBcZV
Thanks to @GoogleDeepMind for inviting me to an early research preview of Genie 3 the other week! I've had a chance to play with it myself and I have some thoughts about what this might mean for video games and game development: https://t.co/CqA4Jla3tO
0
1
7
Overall, I believe Genie 3 is an incredible breakthrough, opening up so many new research directions and applications. In particular, it paves the way for new forms of computer vision/robot training, as well as data generation for foundation models in many domains.
1
1
5
Another difficulty was precisely arranging the world into sections (a jungle on the left, a desert on the right). Instead, the model tended to merge and blend all these concepts to form a consistent and coherent result, but which may not be exactly what you are looking for.
1
1
6