清华赵行在CoRL 2026提出按时间尺度分层机器人智能
摘要
清华大学助理教授、Galaxea联合创始人兼首席科学家赵行在CoRL 2026早期职业主题演讲中提出，机器人智能应按时间尺度而非感知/规划/控制划分：从秒级的System 2推理，到数十赫兹的System 1策略，再到数百赫兹的System 0接触控制。
摘要由模型根据原文撰写，以原文为准。
正文 · 原文
🎓At CoRL 2026, Hang Zhao — Assistant Professor at Tsinghua and co-founder & chief scientist of Galaxea — gave one of this year's Early Career Keynotes.
His argument: the physical world does not pause while a robot computes. Headline numbers look strong, but the bottleneck sits elsewhere: inference latency, not the controller, sets the pace.
So he splits robot intelligence by timescale, not by perception / planning / control: from System 2 reasoning over seconds, to System 1 policies at tens of hertz, down to System 0 for contact at hundreds of hertz. Galaxea fields one model per timescale — Fast-WAM, G0.5, ActionCodec2.
Fast-WAM alone cuts single-step inference from ~800ms to 190ms, a 4x+ speedup. Slides below 👇
Keynote: http://2026.corl.org · Paper: http://arxiv.org/abs/2608.11739
来源：Galaxea Dynamics · x.com
引用原文
GalaxeaDynamics. 🎓At CoRL 2026, Hang Zhao — Assistant Professor at Tsinghua and co-founder & chief scientist of Galaxea — gave one of this year's Early C...[EB/OL]. (2026-11-12)[访问日期]. https://x.com/GalaxeaDynamics/status/2120826411258978803.