←
Chenhao Li 📌 CoRL1,163 posts
Chenhao Li 📌 CoRL avatar

Chenhao Li 📌 CoRL

@breadli428

AI & Robotics Doctoral fellow @ETH_AI_Center | Prev. @GoogleDeepMind, @MIT, @ETH_en, @MPI_IS.

◷ Joined December 2014▧ 260 Photos and videos
Chenhao Li 📌 CoRL avatar
Chenhao Li 📌 CoRL@breadli428·Jun 24

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 📌 CoRL avatar
Chenhao Li 📌 CoRL@breadli428·Nov 13

Many people asked me during #CoRL2026 what I think of SuperMimic, given its similarities to my work on (Uncertainty-Aware) Robotic World Models (RWM/RWM-U), but also the differences in their reception. SuperMimic is not the first work to make model-based RL work on real robots. Nor are its core methodological ideas entirely new, or its tasks fundamentally different from those explored in RWM/RWM-U. However: I think there are several reasons why SuperMimic has resonated so strongly with the community. First, it has a compelling motivation and a strong demonstration of how to lower the computational barriers to training large models such as SONIC. Its problem formulation addresses a critical bottleneck faced by many researchers in a rapidly growing field. The emphasis on improving training efficiency is both timely and practically important. This was a much harder direction to pursue when I was developing RWM/RWM-U, as SONIC had barely emerged or was only just becoming available. At the time, motivating MBRL through online adaptation, as we did in RWM, perhaps appealed to a smaller audience than the growing community interested in SONIC. Second, its technical contribution is impressive. In particular, the policy optimization approach using first-order gradients through learned dynamics is very interesting. I know firsthand how challenging it can be to make BPTT work reliably in this setting, having struggled with it during my own work. I appreciate the technical effort behind this approach. Is SuperMimic overhyped? I would frame the question differently: I was just as excited about the results of RWM/RWM-U a year ago as many people are about SuperMimic today. I believe MBRL deserves much more attention than it has historically received, and I see its growing recognition as a positive development for the entire robotics community. Sometimes, ideas are simply ahead of their time. Neither RWM/RWM-U nor SuperMimic represents the first effort to make MBRL work on real robots. There is a long history of research in this direction, building on foundational work in differentiable simulation and model-based control. I hope we can do a better job of recognizing and acknowledging these earlier contributions. Research visibility and exposure can differ substantially across institutions and communities, but those differences should not prevent us from appreciating the intellectual foundations on which newer work builds. Finally, I want to acknowledge that SuperMimic handled its relationship to RWM and RWM-U very respectfully. The paper, presentation, and code release all provided sufficient and thoughtful acknowledgment of our work. @qiayuanliao also personally reached out to express gratitude for the inspiration he drew from it. As a researcher, it means a lot to see one’s work recognized and treated with respect. Ultimately, I am happy to see MBRL receiving the attention it deserves. I hope this momentum continues, and that we collectively build on and give credit to the many contributions that have brought the field to where it is today. RWM (sites.google.com/view/roboti…) RWM-U (sites.google.com/view/uncert…)

Post attachment

SuperMimic has won the CORL’s best paper award. Thanks for everyone’s support!

Post attachment
Chenhao Li 📌 CoRL avatar
Chenhao Li 📌 CoRL@breadli428·Nov 12

🎤 Tomorrow at Workshop on Modeling Uncertainty in Robotic World Models, I will talk about Uncertainty-Aware Robotic World Model, together with Leslie, Ani @Majumdar_Ani, and Panpan. 📍4:05 PM · Rooms 310–311 · JW Marriott uncertainwm2026.github.io And some thing new:

Post attachment

Hope to see you at our Modeling Uncertainty in Robotic World Models Workshop at CoRL on Nov 12th from 2-6pm! Invited speakers: Leslie Kaelbling (MIT), Anirudha Majumdar @Majumdar_Ani (Princeton / Google DeepMind), Panpan Cai (SJTU), and Chenhao Li @breadli428 (ETH Zurich / Google DeepMind) Website link: uncertainwm2026.github.io

Post attachment
Chenhao Li 📌 CoRL avatar
Chenhao Li 📌 CoRL@breadli428·Nov 11

Happening today! 📍 10:30-12:00 JW Marriott 2nd floor! ➡️ 156 Morphology-Conditioned World Model modanesh.github.io/papers/qw… ➡️ 164 EgoHTR: Egocentric 4D Human Terrain Traversal egohtr.github.io/ ➡️ 174 ADAPT: Online Text-Driven Humanoid Control wuyan01.github.io/ADAPT-proj…

Heading to #CoRL2026 in Austin 🤠 Come say hi! 📌 Poster Session 3 · Wed 11/11, 10:30 AM–12:00 PM · JW Marriott 🐕 Morphology-Cond. World Model 🚶 EgoHTR 🤖 ADAPT 🎤 Workshop talk on (Uncertainty) Robotic World Model · Thu 11/12, 4:05 PM · Rooms 310–311 uncertainwm2026.github.io

Post attachment
Chenhao Li 📌 CoRL avatar
Chenhao Li 📌 CoRL@breadli428·Nov 11

Craziest question heard at #CoRL2026: “Why do you care about delay?” Craziest claim: “Tactile is important but haptic is not” Craziest word: “action-conditioned world model”

Chenhao Li 📌 CoRL avatar
Chenhao Li 📌 CoRL@breadli428·Nov 10

The metric shouldn’t stop at success rate but also efficiency. Moving this slow means everything is almost quasi-static and we know the models today are good at reasoning on images. The improvement to me comes from training domain inclusion rather than better “robotic” skills.

Fable 5.2 shows substantially improved skills in controlling physical robots, even on tasks it has never explicitly been trained to do. On RoboDojo, a benchmark with a simulated two-arm robot, it outperformed other models and most AI systems built to control robots.

0:23
Chenhao Li 📌 CoRL avatar
Chenhao Li 📌 CoRL@breadli428·Nov 5

Heading to #CoRL2026 in Austin 🤠 Come say hi! 📌 Poster Session 3 · Wed 11/11, 10:30 AM–12:00 PM · JW Marriott 🐕 Morphology-Cond. World Model 🚶 EgoHTR 🤖 ADAPT 🎤 Workshop talk on (Uncertainty) Robotic World Model · Thu 11/12, 4:05 PM · Rooms 310–311 uncertainwm2026.github.io

Post attachment
Boris Belousov avatar
⟳ Chenhao Li 📌 CoRL retweeted
Boris Belousov@_bbelousov·Nov 5

How do you make a humanoid follow text commands? NVIDIA's Kimodo, for example, generates a kinematic motion from text, and a separate controller then tracks it. But the generator does not know what the robot can track, so its references can become infeasible when commands switch quickly. ADAPT from ETH Zurich (CoRL 2026) merges the two steps: one diffusion policy maps text straight to joint actions on a Unitree G1 and switches skills mid-motion. How it works • Data: human motions (AMASS) with frame-level text labels (BABEL) are replayed on the G1 by an RL motion tracker in simulation. Successful rollouts become state-action-text training data. • A small causal transformer denoises the next 15 frames of actions and robot states together, given 5 past frames and a CLIP text embedding. A new command just swaps the text embedding. Two denoising steps take about 2 ms, so it runs closed loop at 50 Hz. • Behavior cloning alone falls when skills switch. A residual RL policy corrects only the lower-body actions and is rewarded for staying close to the diffusion model's own predicted next state. Does it pay off? • In simulation, with commands switching every 5-10 s, 98.4% of 20 s rollouts end without a fall, vs 92.3% for the best baseline (LangWBC, end-to-end) and 76.4% for text-to-motion plus tracking, even though the latter sees the whole command schedule in advance. Without the residual policy: 80.4%. Caveats • Success only means no fall, and the residual policy costs some text accuracy: text-motion retrieval R@1 drops from 59.5% to 44.6%. • The real-robot test is 5 trials each for 4 skills (walk and jog 5/5, kick 4/5, jump 3/5). Code is "coming soon". By @Yan_Wu_7, @breadli428, @kaifengzhao, Gen Li, Marco Hutter and @SiyuTang3.

🎯Demo is on the shelf! wuyan01.github.io/ADAPT-proj… Try out the language-conditioned end-to-end humanoid control in your browser! And our paper: 📄 ADAPT: Agile Diffusion Action Priors for Robust and Steerable Online Text-Driven Humanoid Control arxiv.org/abs/2609.00677

0:57
Chenhao Li 📌 CoRL avatar
Chenhao Li 📌 CoRL@breadli428·Nov 5

🎯Demo is on the shelf! wuyan01.github.io/ADAPT-proj… Try out the language-conditioned end-to-end humanoid control in your browser! And our paper: 📄 ADAPT: Agile Diffusion Action Priors for Robust and Steerable Online Text-Driven Humanoid Control arxiv.org/abs/2609.00677

0:57

Dominant text-to-motion pipelines generate kinematic motions for a separate tracker. Instead, in our @corl_conf work ADAPT, we solve humanoid language control end-to-end closed-loop, where the robot must continuously respond to changing commands. 🌐Page: wuyan01.github.io/ADAPT-proj…

0:16
matthias heyrman avatar
⟳ Chenhao Li 📌 CoRL retweeted
matthias heyrman@MatthiasCaro·Nov 4

How do you train a humanoid on sparse terrain inside a world model? Terrain-Grounded Imagination: the model learns only the robot's dynamics. The terrain is a differentiable heightmap lookup at every imagined step, so gaps and edges stay exact. Code: github.com/leggedrobotics/te…

0:41
Manthan Patel avatar
⟳ Chenhao Li 📌 CoRL retweeted
Manthan Patel@patelm99·Nov 4

We gave a robot the walking directions you'd get from a web map. Then we let it find its own way through Zurich: crosswalks, stairs, crowded streets, construction sites, 2.35 km in total. Introducing WayTrace 🧵👇

0:20
Chenhao Li 📌 CoRL avatar
Chenhao Li 📌 CoRL@breadli428·Nov 3

The same diffusion policy can be reused as a steerable text-conditioned motion prior for downstream task adaptation. “run” to goal (green tape) 🆚 “walk while bending over” to goal

0:09
0:20
Chenhao Li 📌 CoRL avatar
Chenhao Li 📌 CoRL@breadli428·Nov 3

ADAPT learns a diffusion-based action prior from text-labeled humanoid trajectories to execute diverse skills from language commands. A lightweight residual RL policy on top of the frozen controller improves long-horizon robustness and enables smooth prompt switching.

Post attachment