TuSimple co-founder Wang Naiyan leaves Xiaomi to found an embodied-AI brain startup
量子位 · wechat · 2026-09-08
Wang Naiyan, L3 technical lead at Xiaomi Auto and co-founder/chief scientist of TuSimple, officially left Xiaomi on September 7 to found a startup focused on the embodied-AI "brain" — core perception, decision-making and planning for robots, possibly including its own robot hardware. His founding team includes several TuSimple veterans, and multiple top-tier VC funds are in close talks.
Background: Wang was the first researcher to apply deep learning to visual tracking (2013), co-developed MXNet, led TuSimple's algorithm team to 10 first-place results on KITTI and Cityscapes within a year, later shipped L2+/L3量产 products, and joined Xiaomi in May 2024 as one of four core leads of its 1,800-person autonomous driving org.
Technical thesis: the embodied-AI field is stuck between VLA (strong semantics, weak physical/causal modeling) and WAM (strong long-horizon planning, huge compute) with no data standard consensus. His DriveZero paper proposes a decouple-and-distill paradigm — RL-trained decision policies in simulation, distilled perception from foundation models, then merged — sidestepping the sim-to-real and data-route chicken-and-egg deadlock. His view: "understanding the world" and "taking actions" should not be fused in one black-box model.
More from Companies & People
- Inside Alpha School's high school boot camp: tweets, heart-rate monitoring and a $50k tuition — benjaminjriley · 2026-09-08
- OpenAI chief scientist Jakub Pachocki says frontier labs should slow down — cccalum · 2026-09-08
- Alpha School reportedly requires high schoolers to gain 100 X followers as boot camp requirement — benjaminjriley · 2026-09-08
- Shovel digs out how work actually gets done before deciding what to automate with AI — jdjohnson · 2026-09-08
- Blogger: Of course OpenAI reads your 'private' Codex logs and stealth-nerfs models — Linahuaa · 2026-09-08
- Teknium reveals the Hermes Agent team is just 3 people, plus 1 on GUI — Teknium · 2026-09-08