Yunhao Cao

Learning to make robots learn @ Cornell

[Yunhao Cao @QuantummCookie]

@NoriRobotics Curious if there's a low-level SDK for this? Might also be a good platform for research

[Qianxu Wang @helloqxwang]

Introducing KITE 🪁, a paradigm that transfers the interaction intent instead of joint-space actions across structurally different manipulators. Deploying on a new robot needs only its kinematics model (URDF), no new task demonstrations. 🧵1/8

[Yunhao Cao @QuantummCookie]

Check out our paper and project page for more details!
🌐 Website: https://t.co/wWfy2O07Il
📝 Paper: https://t.co/Ob7eqK9dPU

Huge thanks to my collaborators Shitong Liu, @chaof1234 , Meryl Zhang, @XuanchenLu , and mentors @andrewhowens , @KuanFang .

[Yunhao Cao @QuantummCookie]

Introducing Unified Motion-Action (UMA) Model, a robot foundation model that uses 3D object motion as a shared interface for heterogeneous robot learning. UMA treats motion and action as co-evolving variables, enabling knowledge transfer across data sources and versatile inference. 🧵 1/n

[Yunhao Cao @QuantummCookie]

Our analysis isolates the design choices that matter, confirms that UMA scales with both data and model size, and traces its remaining errors to execution rather than perception. 10/n

[Yunhao Cao @QuantummCookie]

@SeeLos @thdxr Interesting...Thanks for the info

[Yunhao Cao @QuantummCookie]

@sama GPT 5.5 today?

[Yunhao Cao @QuantummCookie]

@kevin_zakka Cool!

[Yunhao Cao @QuantummCookie]

CorDex utilizes 3D contact keypoints to transfer grasping knowledge from single human demonstrations to hundreds of synthetic objects. Check out our work on ArXiv!

[Xingyi He @XingyiHe1]

🚀Excited to share our #ICRA2026 work, CorDex!

CorDex learns functional dexterous grasping from a single human demo, generalizes to unseen objects using a single-view observation, and enables robots to use tools for long-horizon tasks.

Project Page: https://t.co/LAkO8LkhK5

[Jay Karhade @JayKarhade]

Introducing Any4D, a unified transformer for fully feed-forward, dense, metric-scale 4D reconstruction from flexible inputs!

Any4D regresses per-pixel motion + geometry across frames in one pass — 15× faster, 2–3× more accurate reconstructions ⚡📈

Details + code below 👇

Exciting collab with @Nik__V__ @YuchenZhan54250 Tanisha Gupta @akashshrm02 @smash0190 @RamananDeva

[Yunhao Cao @QuantummCookie]

@codewithimanshu Thanks Himanshu!

[Yunhao Cao @QuantummCookie]

Introducing Correspondence-Oriented Imitation Learning (COIL), a self-supervised framework for robotic manipulation using 3D keypoint trajectories as a flexible, unified task representation!