Ex-Huawei AI chief Tang Duyu startup builds proactive physical AI that acts before you ask
量子位 · wechat · 2026-10-15
Tang Duyu, former Huawei AI chief expert and head of the Xiaoyi multimodal model, has founded Guanxing Digital to build proactive physical AI — systems that continuously read the event stream of the real world and decide autonomously when and how to intervene, instead of waiting for human commands.
The team designed a dual-model architecture, ProAgent: a lightweight StreamingJev acts as a sentry that cheaply decides whether intervention is needed, while a ResponseModel plans the response and picks which device to invoke. For training data, they automatically extract decision sequences with reasoning from public videos, teaching the model to distinguish "something changed" from "a need has formed." They also built the ProactiveXBench benchmark (Timing / Progress / Outcome, targeting the three classic failure modes), report beating GPT-6-class general models on intervention-timing metrics, and have deployed the system on a Reachy Mini robot across 60 guidance/alert/interaction cases. Tang argues proactive capability won't emerge from scaling alone — it requires separate work on risk judgment, responsibility boundaries and timing.
More from Embodied
- AI concept: Pokémon Snap on a foldable iPhone with dual-screen camera UI — yungcontent · 2026-10-23
- HumanoidPose recovers humanoid robots' full-body joint states from RGB video alone — cholz · 2026-10-23
- Peeling the iPhone Duo's nano-texture film reveals a surprisingly uneven screen underneath — film_girl · 2026-10-23
- ALPS: a single visual encoder unifying state, language, and pixels for world models — cihangxie · 2026-10-23
- XPeng IRON does voice-free handshakes, re-inferring motion every 0.2s via VLA 2.0 — CyberRobooo · 2026-10-23
- Viral humanoid robot cucumber-cutting video debunked: 2x background speed exaggerates the shaking — jonstephens85 · 2026-10-23