Chenhao Li

Student Researcher @GoogleDeepMind | Embodied intelligence and robot learning | Doctoral fellow @ETH_AI_Center, @leggedrobotics | Prev. @MIT, @ETH_en, @MPI_IS.

[Chenhao Li @breadli428]

🌎We learn robot control policies from world models with real deployment.

1⃣ Robotic World Model (RWM) corrects errors through online interaction.
🔗https://t.co/MHG6wKijLk

2⃣ Uncertainty-Aware RWM adds uncertainty penalties during policy optimization.
🔗https://t.co/OT0E93N1nM

[Chenhao Li @breadli428]

@SciRobotics It’s a humanoid

[Eric Rosen @_ericrosen]

😍I love the usage of the arms to balance the climb!

More contact means making use of all parts of the humanoid!

[Takahiro Miki @ki_ki_ki1]

Cool work done by my students @ctki49 @KehanWen170077
Perceptive motion generator + tracker on rough terrain.

[Chenhao Li @breadli428]

@jack_polloway The techniques are quite different tho. The dog operated on a fixed set of skills while the humanoid needs adaptive references to be generated on the fly.

[Chenhao Li @breadli428]

❗️Flat-terrain tracking is solved. Rough terrain breaks everything, because the reference itself has to change. So we built a system that generates references as it goes.

🦿Parkour over boxes, hurdles, stairs - all onboard.

🔗 https://t.co/ZmgNu1e72h
📄 https://t.co/sdksBQfDa6

[Chenhao Li @breadli428]

@DevRajyagu18621 Nope, that’s the bias from the initial reference dataset

[Junzhe (JJ) He @JayHe748646]

Excited to share our recent work on whole-body humanoid locomotion for challenging terrain traversal!

Diffusion-based planner + RL WBC = general purpose locomotion controller

Led by @ctki49 @mxu_cg @KehanWen170077 at @leggedrobotics and @xbpeng4.

[Chenhao Li @breadli428]

⛰️We try to push motion learning beyond what one can do on flat ground.

Kudos to Zewei Zhang @ctki49, Kehan Wen @KehanWen170077, and Michael Xu @mxu_cg, who made this a reality.

Check out now
https://t.co/ZmgNu1e72h

[C. Zhang @ChongZzZhang]

https://t.co/531oWVobqF

In this work we explored how to repurpose terrain-aware human locomotion skills in a scalable and versatile way: like the action-free world model, we predict how humans will move online, and then let the humanoid track it with RL.

[Chenhao Li @breadli428]

This work came from a fruitful collaboration between @SFU and @leggedrobotics @ETH_en, supported by Michael Xu @mxu_cg, Junzhe He @JayHe748646, Chenhao Li @breadli428, Takahiro Miki @ki_ki_ki1, Clemens Schwarke, Chong Zhang @ChongZzZhang, Xue Bin Peng @xbpeng4, Marco Hutter.

[Chenhao Li @breadli428]

Watching my students Zewei @ctki49 and Kehan @KehanWen170077 grow into independent researchers and experts in humanoid locomotion has been the most rewarding part of this project — every bit as gratifying as the technical results. Couldn't be prouder of them. 👏

[Chenhao Li @breadli428]

@dan_in_robots Why do I see it has all head, torso and limbs?

[Chenhao Li @breadli428]

I can’t find a reason why I can’t call it a humanoid.

[Boyuan Chen @Boyuan__Chen]

What if a robot could move, recover, sense, and interact with the world equally well in any direction?

Today, I’m incredibly excited to share our new paper on computational robot design published in Science Robotics!

This project started with a simple yet deep question that kept coming back to us: What if symmetry in robots was not just about appearance, but about how they can dynamically interact with the world?

In this work, we introduce the idea of dynamic symmetry: designing robots with nearly uniform dynamic actuation capability in all directions. We formalize this through a new theoretical measure called dynamic isotropy, and show that pushing robots toward extreme dynamic symmetry unlocks new capabilities in mobility, robustness, resilience, and multifunctionality.

To explore this idea, we built Argus, a family of spherical robots with radially distributed actuators and omnidirectional sensing. Watching Argus come alive for the first time was honestly one of the most exciting moments for our lab.

Our 20-leg physical robot can:
- Move omnidirectionally without needing to reorient first
- Traverse cluttered and deformable terrains
- Recover from disturbances and actuator failures
- Carry heavy payloads
- Climb between walls under lunar gravity
- Perform whole-body loco-manipulation while continuously sensing the environment

What excites me most is that this work is not only about one robot. It explores a broader idea of embodied intelligence: Can symmetry become a fundamental design principle for building robots that are more adaptable, resilient, and capable of interacting with complex real-world environments?

- Website (paper, code): https://t.co/g3dzMsKk55
- Video: https://t.co/X9oTSCue3q

[Chenhao Li @breadli428]

@gazorp5 I’m serious.

[Chenhao Li @breadli428]

@CSProfKGD @oier_mees @mapo1 The efforts to put all robots in the same hall paid off.

[Chenhao Li @breadli428]

@realyunlong @ETH Many don’t even know ETH 😢

[Chenhao Li @breadli428]

@isgeles @ylecun @davsca1 Did he say you should use JEPA to train your multi-agent MBRL?

[Chenhao Li @breadli428]

@abhishekunique7 I believe in Robotic World Model, the model is designed to be finetuned during online adaptations, then the policy is further finetuned within this model.

There’s an offline version of it, Uncertain Aware Robotic World Model. You might have been referring to that one. 🤔

[Chenhao Li @breadli428]

@KKawaharazuka Legend

[Chenhao Li @breadli428]

@KKawaharazuka Let’s goooo!