Make the physical world iterate as fast as software.

Hi 👋 I'm Yuxi. Software improves in a tight loop: write, test, ship, learn. The physical world has never had one this fast. I'm building physical intelligence — AI that designs, simulates, builds and learns, with robots as its hands — to close that loop for engineering, manufacturing and science, and make each turn faster than the last.

Chat with Yuxi

Previously: Google DeepMind · ByteDance Seed
Known for: SpatialTracker · CoDeF · 7k+ GitHub stars

Looking over London and the Thames at sunset

Looking up at a telescope dome on Mauna Kea, above the clouds

The other sideSome loops I'd never speed up:
the sun going down, a telescope waiting for dark.

Selected work

7k+ GitHub stars · 3 CVPR highlights · Best Paper, ICLR 2026 ES-Reasoning Workshop

CVPR 2026 · Highlight

SpatialTree

A four-level hierarchy of spatial abilities in multimodal LLMs, and its benchmark.

Project Paper Code

ICCV 2025

SpatialTrackerV2

Feed-forward 3D point tracking: tracks, depth and camera pose in one model.

Project Paper Code Demo

CVPR 2024 · Highlight

SpatialTracker

Tracks any 2D pixel in 3D space, where motion is simpler than in 2D.

Project Paper Code

CVPR 2024 · Highlight

CoDeF

Video as a canonical image plus a deformation field. Co-first author.

Project Paper Code

Experience

NowFounder, stealth startup
2026Google DeepMind, Student Researcher
2025–26ByteDance Seed, Top Seed research intern
2023–25Ant Group, Research intern
2023–27Zhejiang University, PhD
2019–23Wuhan University

Updates

2026Started a company. Can't say more yet.
2026SpatialTree: CVPR highlight, and Best Paper at the ICLR ES-Reasoning Workshop.
2025SpatialTrackerV2 accepted to ICCV.
2024SpatialTracker and CoDeF selected as CVPR highlights.