清华系初创厘清智能完成数亿元种子轮融资，构建全栈物理AI基础设施
英文摘要
Li Qing Intelligent (厘清智能), a Physical AI startup founded in April 2026 by Tsinghua professor and former Nvidia researcher Li Yiming, has raised hundreds of millions of RMB in a seed round from investors including Shunwei Capital, Sequoia China, Hillhouse Venture Capital, Frees Fund, Star Connect Capital, and industrial backers like Zhiyuan Robot. The company builds a full-stack infrastructure covering self-developed data collection gloves (reducing per-unit cost to RMB level), a differentiable physics engine for Real-to-Sim-to-Real reinforcement learning, and a world model integrated across pre-training and post-training. Its system already achieves fine manipulation skills (cutting, screwing, inserting, stirring) across different grippers and robotic arms, targeting B-end scenarios from manufacturing to retail. The team plans to release a cross-industry world model by end of 2026 and scale deployments by 2028, aiming to deliver a hardware-agnostic, end-to-end solution like an “iOS for physical tasks.” The 50-member team, with an average age of 23, emphasizes full-stack soft/hard integration rather than being a “world model company.”
中文摘要
清华系初创「厘清智能」于2026年4月成立，2个月内即完成数亿元种子轮融资，投资方包括顺为资本、红杉中国、高瓴创投、峰瑞资本、星连资本及智元机器人等产业资本。公司由前英伟达研究员、清华助理教授李一鸣带队，构建自研数据管线（成本压至人民币级别的手套，规模达百万小时）、可微物理引擎（实现Real-to-Sim-Real闭环）和世界模型的一整套物理AI基础设施。系统已能实现切割、旋拧、插拔等精细操作，可跨灵巧手和机械臂部署于制造、零售、酒店等场景。团队计划2026年底发布跨B端场景的世界模型，2028年实现方案规模化落地，最终交付软硬一体、跨本体跨场景的通用物理AI基础设施。团队平均年龄23岁，强调全栈软硬一体而非单纯的世界模型标签。
关键要点
Seed round of hundreds of millions of RMB, closed within two months of founding, backed by top VCs and industrial investors like Zhiyuan Robot.
成立两个月内完成数亿元种子轮融资，投资方包括顺为、红杉中国、高瓴等顶级机构及智元机器人等产业资本。
Full-stack Physical AI infrastructure: self-developed tactile glove for scalable data collection, differentiable physics engine for Real-to-Sim-to-Real, and world model integrated into pre-training and RL post-training.
全栈物理AI基础设施：自研低成本数据采集手套、可微物理引擎实现Real-to-Sim-Real、世界模型贯穿预训练和强化后训练。
System already performs fine manipulation skills (cutting, screwing, inserting) and deploys across different embodiments (grippers, arms) for B-end scenarios like manufacturing and retail.
系统已实现切割、旋拧、插拔等精细操作技能，能在不同灵巧手和机械臂之间跨形态部署，适用制造、零售等B端场景。
Roadmap: cross-B-scenario world model release by end of 2026; scaled solutions by 2028, with the goal of a hardware-agnostic universal platform analogous to iOS for physical tasks.
规划：2026年底发布跨B端场景的世界模型，2028年实现方案规模化落地，终极目标是成为像iOS一样的通用物理操作基础平台。
The 50-person team (average age 23) emphasizes full-stack soft/hard integration and talent scarcity, positioning itself not as a model company but a Physical AI Infra provider.
50余人团队平均年龄23岁，强调软硬一体人才稀缺，定位为物理AI基础设施公司而非单一模型公司。