Jianfei Yang

🎵 Today I don't feel like doing anything… but I'm finally doing what I actually want to do.

Microsoft services went down today.
No Outlook. No emails. No notifications.
For the first time in a long while, I received ZERO emails. 😂
And something magical happened.
I had meaningful discussions in meetings.
I read papers without interruptions.
I sat down and actually thought about research.
No urgent replies. No endless administrative tasks. No inbox anxiety.
Just science, ideas, and curiosity.

I almost forgot this is what being a researcher is supposed to feel like.

Maybe the biggest creativity breakthrough of 2026 isn't another AI agent.
It's Microsoft Outlook going offline. 🤣

🤖 How can we make humanoid robots walk quietly?

Humans can walk softly, but humanoid robots often stomp loudly. Why?

Because walking isn't just about moving your legs—it's about controlling how your feet interact with the ground.

At #CoRL 2026, we introduce ADAPT, using physics-guided RL to help humanoids walk with lighter, quieter footsteps while adapting to unexpected pushes and payloads.

👣 Less stomping, smoother walking!

🎥 Project & Videos: blyu413.github.io/adapt-locomoti…

📄 Paper: arxiv.org/abs/2606.16542

#HumanoidRobotics #PhysicalAI

Very excited to see that ODE policy + simple pi0.5 or WAM fine-tuning achieved great results on variants of LIBERO, tested by the @huggingface team.

Furthermore, ODE exhibits faster training and smoother trajectories without any RTC.

Feel free to read and try ODE policy at ntumars.github.io/ODE/

🚀 Nature does not move in discrete jumps. Why should robot policies?

Super excited to release ODE Policy — perhaps our most exciting breakthrough in my lab this year! 🤖🔥

Think about how humans reach for a cup. Our movements evolve continuously, governed by underlying dynamics. Yet, most robot policies — from Diffusion and Flow Matching to VLAs — learn to predict explicit actions.

What if robots learned how actions evolve, instead of predicting actions?

💡 That's ODE Policy for Embodied AI.

We use a neural Ordinary Differential Equation (ODE) to learn continuous action dynamics, rather than directly generating actions.

One simple idea. Five types of robots. Multiple learning paradigms.

✅ Single-arm manipulation
✅ Dual-arm manipulation
✅ Dexterous hands
✅ Humanoid whole-body control
✅ Aerial robots

And it goes far beyond imitation learning:

⚡ VLA (π0.5): 96.0% vs. 75.5% success after one epoch.
🌍 4.5B World Action Model: 82.6% vs. 24.6% success, with ~4× faster action generation.
🏃 RL: More efficient humanoid whole-body motion learning.

For years, we've focused on scaling robot data, models, and compute.

But what if the next fundamental breakthrough comes from rethinking WHAT robots learn, rather than just HOW MUCH they learn?

Huge congratulations to Jindou Jia and our amazing team! 👏

📄 Paper: arxiv.org/abs/2611.01729
🌐 Project: ntumars.github.io/ODE/

#ODEPolicy #EmbodiedAI #PhysicalAI #RobotLearning #WorldModels #VLA #MARSLab

🤖 Introducing ω-0 (OMEGA-0) — a whole-body World Action Model for humanoids.

Can one model make a humanoid walk, manipulate, and coordinate its whole body simultaneously across many real-world tasks?

ω-0 does exactly that.

⚡ One unified model for multi-task whole-body loco-manipulation
🧠 Predicts future visual latents while generating executable whole-body actions
🏠 Trained with ω-HOME: 40+ hours, 24 household tasks, 4,827 episodes
🔥 81.8% real-world success rate across 11 evaluation tasks

No separate locomotion/manipulation policies. One model. Whole body. Multiple tasks.
Humanoid intelligence should not be a collection of isolated skills. It should be a unified whole-body policy.

🔗 gentlefress.github.io/OMEGA-0_page/
📜 arxiv.org/abs/2608.06375

One-take, uncut real-world demo👇