@breadli428
Account based inUnited Kingdom!
About this account
- Account based in
- United Kingdom
- Connected via
- United States App Store
X says this location may be affected by a proxy or VPN.
Account-level information from X, not a live location or the device used for a specific post.
AI & Robotics @GoogleDeepMind | Doctoral fellow @ETH_AI_Center, @leggedrobotics | Prev. @MIT, @ETH_en, @MPI_IS.
Zurich, Switzerland
Joined December 2014
- Tweets1K
- Following475
- Followers7.8K
- Likes2.1K
Pinned Tweet
I decided to review and explicitly post about the limitations of all my previous papers because I believe this is the fundamental driving force behind research, especially in this era of PRs and bubbles.
Chenhao Li retweeted
Today, we are leaving stealth mode at Vesoma: vesoma.com/news/introducing-…
Our vision: We want to build and ship the most versatile and trustworthy humanoid to give AI the body it needs to give us a helping hand and unblock the bottleneck of physical labour.
🧵1/6
Chenhao Li retweeted
Met some absolute legends of the simulation community, an incredible conference hosted by nvidia and google in london!
Working in the heart of London is a privilege. 🎡
A few photos of the skyline from my office @GoogleDeepMind 🏙️
Hard to get bored with it.
We are releasing a dataset with human motions together with terrain reference collected both indoors and outdoors.
Now, contact signals can be extracted to help with learning terrain interactions.
Presenting in Nov 2026 @corl_conf!
Try out EgoHTR now!
🔗 egohtr.github.io/
We are releasing EgoHTR, a dataset with both human motions and terrain references, accepted @corl_conf.
📖 Paper: lnkd.in/eeepUefs
🌐 Project Page: egohtr.github.io
• 55 scene-aligned sequences • 150k+ frames • rough-terrain environments • multi-modal 3D scene
It’s great to see more people openly discussing the true limitations of their own work.
No one understands the strengths and weaknesses of your method better than you do. That puts you in the best position to contribute to principled research—and to help others build on it.
🧵
🎉 FADA is now accepted to CoRL 2026!
Along with the code, we’re releasing something a little different: a timeline of the research journey behind FADA — the ideas we tried, what failed, the choices we made, and the lessons we learned along the way.
lecar-lab.github.io/FADA-hum…
Research papers are great at telling us what worked. They are much worse at telling us everything researchers tried before getting there.
We think exposing that trajectory can give the community a much richer and more honest ground-truth context particularly in the age of AI-assisted robotics research.
We’d love to see more projects share their research journeys this way: turn your update slides to a shareable timeline!!.
Understanding why certain choices were made (and which mistakes didn’t work) is just as valuable as the final result.
Inspired from @breadli428's efforts on adding limitation threads to his works and @TairanHe99's timeline videos in VIRAL.
🧵
Chenhao Li retweeted
A few moments from IWIALS 2026 in Kleinwalsertal!
Thanks to @m_wulfmeier, Jesús Tordesillas, and @breadli428 for their talks; to everyone who joined us for a week of research fun; and to @Jan_R_Peters, @mdrolet01, and @erikhelmut for organizing!
Was amazing to learn how much has been achieved already at @nomagicAI with @m_wulfmeier! Really enjoyed the time catching up with old and new folks at the workshop!
Deeply enjoyed talking about mastery-first physical AI at the IWIALS workshop! Old friends and many new ones - looking forward to how this group will continue to shape European AI.
On mastery: we've made strides on generality and now need to demonstrate real-world more than ever before to justify funding. Even generalisation to 99% is only reducing per-task budget by 50% following Chinchilla scaling (which is a bit pessimistic but even optimistic versions leave 30%). New recipes, data sources, and more is needed!
Thanks to the invitation to @Jan_R_Peters @mdrolet01 @erikhelmut! Hope we get to catch up soon @breadli428 @RLioutikov! We're continuing to hire masterful RSs and REs @nomagicAI!
Surprising how easily one can prompt the agent to train RL locomotion policies in simulators now!
One prompt 💬. One simulator. A trained robot policy 🤖.
🚢 HARBOR autonomously builds the task, designs rewards, trains, tunes, and evaluates — end to end. Now accepted at #CoRL2026.
From prompt to policy, fully autonomous, 1.5h. 👇
Chenhao Li retweeted
Two new Gemini models are here to help scale your AI agents and secure code:
🔘 3.8 Flash: our most intelligent model yet with significant gains from 3.7 Flash across software engineering, agentic tasks, and multi-step reasoning.
🔘 3.8 Flash Cyber: our most capable cybersecurity model with frontier-level vulnerability detection and automated patching.
🎯 Limitations:
Similar to VLAs, residual RL may unlearn/overwrite skills of the base policy and collapse its diversity. For specific tasks, using a complex hierarchical structure with general base policy needs to be justified.
📢 Excited to share our latest work @corl_conf: Motion Priors Reimagined: Adapting Flat-Terrain Skills for Complex Quadruped Mobility! #CoRL2025
🐶Now ANYmal learns to walk (hop) from real dog motions!
📄Paper: arxiv.org/abs/2505.16084
🌐Project Website: anymalprior.github.io/
Great honor and fun to talk about robotic world models at International Workshop of Intelligent Autonomous Learning Systems @ias_tudarmstadt.
First time as a speaker invited by @Jan_R_Peters with Jesus and @m_wulfmeier.
Excited to learn talented people around in robot learning!
Mind blowing 🤯
This is killing research in many academic labs. Nor do I believe any published research was ever close to this, regardless of the posted videos.
Does it still make sense to work on sim-to-real locomotion?
What’s more impressive than the speed is the steering.
Running straight for 100m can be really hard.
