Robotic Systems Lab

The Robotic Systems Lab designs machines, creates actuation principles, and builds up control technologies for autonomous operation in challenging environments.

[Robotic Systems Lab @leggedrobotics]

Our new work WayTrace: an off-the-shelf VLM, an onboard trace model, and an RL policy take a robot from web-map directions to 2.35 km of real-world missions in Zurich.

Details in the thread 👇

[Manthan Patel ✈️ CoRL @patelm99]

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 🧵👇

[Robotic Systems Lab @leggedrobotics]

This work is led by Yan Wu @Yan_Wu_7 and Chenhao Li @breadli428, with the team Kaifeng Zhao, Gen Li, Marco Hutter @leggedrobotics and Siyu Tang @SiyuTang3.

📄 Paper: https://t.co/TKb65ujdrk
🌐 Project page: https://t.co/5lJdGnaIN3
📍#CoRL2026 Session 3, Poster #174!

[Robotic Systems Lab @leggedrobotics]

In our latest work @corl_conf, we present ADAPT: an end-to-end framework for interactive, text-conditioned humanoid control, in contrast to dominant pipelines that generate kinematic motions for a separate RL tracker. 🤖

🌐 https://t.co/5lJdGnaIN3

Full video in🧵

[Chenhao Li @ CoRL @breadli428]

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: https://t.co/87y3FVjT4H

[Robotic Systems Lab @leggedrobotics]

We hope to change ideas on how perception should integrated in rough terrain locomotion with our latest work

Watch Your Step: Anticipated Perception for Humanoid Locomotion

We predict and perceive only where it matters.

🌐 https://t.co/lgDDkoJ3BR
📷 https://t.co/qYZ8QZKFN0

[Chenhao Li @ CoRL @breadli428]

We only care about where our feet are about to land when we walk. Why shouldn’t humanoids?

Introducing Anticipated Perception for Humanoid Locomotion, anticipating upcoming footholds and perceiving only the terrain patches that matter for next step.👣

🌐 https://t.co/hDXr3GUlol

[Robotic Systems Lab @leggedrobotics]

Try the policy in your browser!

Authors:
Matthias Heyrman @MatthiasCaro, Chenhao Li @breadli428, Ioannis Dadiotis, and Marco Hutter

@leggedrobotics @ETH_en @ETH_AI_Center

[Robotic Systems Lab @leggedrobotics]

We walk without worrying about what’s behind us. Why should humanoids?

We introduce Anticipated Perception for Humanoid Locomotion: a new approach that focuses terrain perception on where the feet are about to land. 👣

🌐 https://t.co/lgDDkoJBrp 📷 https://t.co/qYZ8QZLdCy

[Robotic Systems Lab @leggedrobotics]

Introducing DeFM: a foundation model for depth images, built for robotics.

Weights, code and a browser demo are out!

[Manthan Patel ✈️ CoRL @patelm99]

Vision foundation models like DINO and SigLip now power much of robot learning.

But they are built for RGB. What about depth?

Introducing DeFM: a foundation model for depth images, built for robotics.

🌐 https://t.co/KzloKyRZyF

Weights, code, and demo are out today 🧵

[Robotic Systems Lab @leggedrobotics]

Can a humanoid learn faster and better by seeing less?

We train a small model to predict the policy's next steps, and show the policy only the terrain there. Together, they learn to cross gaps, stairs, drops and boxes.

🌐 https://t.co/lgDDkoJBrp
📄 https://t.co/qYZ8QZLdCy

[matthias heyrman @MatthiasCaro]

What does a humanoid need to look at to cross challenging terrain?

People only look around their next few planned steps. So we showed a humanoid only that.

Watch Your Step: Anticipated Perception for Humanoid Locomotion

🌐 https://t.co/W4NNjxI0tX
📄 https://t.co/wDOEsxkYdj

[Robotic Systems Lab @leggedrobotics]

We will see you in Austin, Texas, US, in November at CoRL 2026 @corl_conf!

[Robotic Systems Lab @leggedrobotics]

This work is led by Mohamad Danesh @mo_danesh, together with Chenhao Li @breadli428, Amin Abyaneh, Anas Houssaini, Kirsty Ellis, Glen Berseth @GlenBerseth, Marco Hutter, and Hsiu-Chin Lin.

@mcgillu
@Mila_Quebec
@UMontreal
@ETH_en
@ETH_AI_Center
@leggedrobotics

[Robotic Systems Lab @leggedrobotics]

In our latest work at @corl_conf, we introduce Quadrupedal World Model: a single morphology-conditioned dynamics model, where policies are trained entirely in imagination.

We found generalization emerges by conditioning the dynamics, not the policy.

🔗https://t.co/6Ptc7z7LQP

[Mohamad H. Danesh @mo_danesh]

What if one world model could control many different robots? 🤖🌍

That's QWM.

In this 2 minutes video we show keypoints about QWM: learning across heterogeneous quadrupeds, and transferring zero-shot to unseen morphologies.

No fine-tuning. No warm-up. No retraining.

Let's talk it further at CoRL 🥳

[Robotic Systems Lab @leggedrobotics]

Scaling Rough Terrain Locomotion With Automatic Curriculum Reinforcement Learning

Authors: Ziming Li @ZimingLi97, Chenhao Li @breadli428, Marco Hutter

Published in: IEEE Robotics and Automation Letters ( Volume: 11, Issue: 8, August 2026)

https://t.co/RF93oRGsrd

[Robotic Systems Lab @leggedrobotics]

LP-ACRL removes the need to predefine task difficulty. It measures task-wise learning progress from episodic-reward changes, then uses a softmax to shift sampling toward the current learning frontier, where tasks show stronger positive competence gains at that stage.

[Robotic Systems Lab @leggedrobotics]

After teacher–student distillation, a single policy transfers to ANYmal D and achieves:

• 3.0 m/s on flat terrain
• 2.5 m/s on challenging terrain
• Up to 3.0 rad/s angular velocity across all tested terrains

[Robotic Systems Lab @leggedrobotics]

Real-world locomotion does not follow a single easy-to-hard axis. Terrain type and geometry compound with linear and angular velocity specifications.

In our scaled setup, their Cartesian product forms 600 tasks, making a manually ordered curriculum impossible.

[Robotic Systems Lab @leggedrobotics]

We achieved high-speed rough-terrain locomotion on ANYmal with an automatic curriculum.

LP-ACRL automatically samples terrain types, levels, and velocity commands at the correct time based on policy performance, without predefining an order.

🔗https://t.co/ip3pQ0z2qF
🧵RAL 2026

[Robotic Systems Lab @leggedrobotics]

Congrats to Kaixian Qu, Guowei Lan, René Zurbrügg, Changan Chen, Christopher Mower, Haitham Bou-Ammar & Marco Hutter. Say hi at IROS 2026!

https://t.co/4APFhtIG6t
https://t.co/1tuEgO0cXH
https://t.co/UKPpXcG8i0
https://t.co/byNWeCEce1

[Robotic Systems Lab @leggedrobotics]

VLMs don't understand a robot's body. How can a robot learn from its failures without retraining?

PragmaBot: online in-context learning from real-world experience. The robot reflects on failures, stores lessons in memory, and retrieves them for new tasks.

RA-L · #IROS2026

[Robotic Systems Lab @leggedrobotics]

How much memory belongs in the context? 5 random memories: 17% first-action accuracy on unseen tasks. Entire memory in the prompt: 74%, at 7.5x the tokens. Top-k relevant via RAG: 89%. Relevant context beats more context. As memory grows, what should a robot forget?