Yuxiang Gao // roboticist · founder & CEO, COCO Matrix // 机器人研究者 · COCO Matrix 创始人 & CEO
I teach robots how to learn.
我教机器人，
学会学习。
I'm a roboticist. In April 2026 I founded COCO Matrix, where we build robots that learn the way people do. A few demonstrations, a little language, an occasional correction, and the robot has a new skill.
Robots have kept me busy for ten years, from human-robot interaction research at Johns Hopkins to humanoid world models at Fourier. The longer I built them, the harder one question pressed. Where should robot intelligence go? Scenes change; the person stays. Today's robots compete on how much they know. The ones I want to build will compete on how fast they learn, and people are their best teachers.
我是一名机器人研究者。2026 年 4 月，我创办了 COCO Matrix。我们让机器人像人一样学习，几次示范，几句指导，偶尔纠正一下，它就学会一个新技能。
过去十年，驱动我的一直是机器人。在约翰·霍普金斯研究人机交互，在傅利叶做人形机器人世界模型。造得越久，有个问题就越挥之不去。机器人智能最终应该走向哪里？场景一直在变，人始终没变。今天的机器人，比的是谁会得更多。我想造的机器人，比的是谁学得更快，而人，就是它最好的老师。
// 00 · playground // 00 · 游乐场
A little game. Draw one demonstration and lead the robot to its charger. 来个小游戏，画一次示范，带小机器人找到充电桩。
// draw a route around the boxes, end at the charger. // 绕开箱子，一笔画到充电桩。
// 01COCO Matrix
COCO Matrix is an embodied-model company chasing physical AGI through in-context learning. We train robots to learn new tasks. Everything a robot experiences, the language it hears, what it sees and touches, its own action history, becomes context for the model, and the model adapts on the spot. No retraining, and not a single weight updated. Teaching it takes no robotics knowledge. Do the task once in front of it, tell it what went wrong, and that's enough. Grandparents can teach it; so can a shop manager or a nurse. You teach it the way you'd teach a person, and like a capable apprentice, it learns from one showing. That's the robot we're building.
The name spells out the relationship we want with machines. Coexist, cowork, collaborate, coevolve.
COCO Matrix 是一家具身模型公司，用上下文学习（in-context learning）去做物理 AGI。我们训练机器人学会新任务。机器人经历的一切，听到的语言、看到的画面、指尖的触觉、自己的动作历史，都会成为模型的上下文，模型当场适应，既不用重新训练，也不更新任何参数。教它的人不需要任何机器人知识，做一遍给它看，说一句哪里不对，就够了。爷爷奶奶能教，店长和护士也能教。像教人一样教机器人，聪明的学徒，教一遍就会。我们想造的就是这样的机器人。
名字拼出来的，就是我们想要的人机关系。共存 Coexist、共事 Cowork、协作 Collaborate、共同进化 Coevolve。
// 02A roboticist's training log一个机器人研究者的训练日志
- 2013–17 // pretraining Xi'an Jiaotong University, Honors Youth Program (少年班). B.S. in electronic information, first research running visual SLAM on an autonomous car. 西安交通大学少年班，电子信息专业本科。第一段研究，是给一辆自动驾驶汽车做视觉 SLAM。
- 2017–19 // imitation Johns Hopkins, M.S.E. in Robotics. Built PATI (IUI '19), a projected tabletop where anyone can hand a robot arm a task with a few bare-hand gestures. Watching people teach a machine the way they'd teach a colleague planted the question COCO Matrix now works on. Also led JHU's first RoboMaster AI Challenge team into the DJI finals at ICRA 2018. 约翰·霍普金斯大学，机器人硕士。做了 PATI（IUI '19），一个投影桌面系统，普通人徒手比划几下，就能给机械臂布置任务。看着人像教同事一样教机器，COCO Matrix 要回答的问题，就是从那时候种下的。也带着 JHU 第一支 RoboMaster AI 挑战赛队伍，打进了 ICRA 2018 的 DJI 决赛。
- 2019–22 // exploration Johns Hopkins, Intuitive Computing Lab. Human-robot interaction research. Finished the M.S.E. in CS, left the PhD to go build. 约翰·霍普金斯大学，Intuitive Computing Lab。研究人机交互。硕士读完，从博士退学，去动手造东西。
- 2022–24 // deployment CTO of a healthcare-AI startup, teaching computer vision to read human bodies and movement while guiding real patients through physiotherapy. Understanding people, in a different setting. 一家医疗 AI 公司的 CTO，用计算机视觉读懂人的身体和动作，带真实患者做康复训练。理解人这门功课，换了个场景继续。
- 2024–26 // world models Fourier. Architected the teleoperation backbone behind large-scale humanoid data collection, which at scale is simply people showing robots how to work. Led world-model action policy research on a full-size humanoid, the kind of model base in-context learning stands on. Somewhere in there, COCO Matrix turned from a question into a plan. 傅利叶。搭建了大规模人形机器人数据采集背后的遥操作系统，这件事本身，就是人成规模地做给机器人看。在全尺寸人形机器人上主导世界模型动作策略的研发，机器人要做上下文学习，靠的正是这种模型底座。做着做着，COCO Matrix 从一个问题变成了一份计划。
- 2026– // in-context learning COCO Matrix, founder & CEO. Ten years of watching people teach robots, gathered into one company. We build embodied models that learn in context, from demonstration, language, and touch, for the people they will serve. COCO Matrix，创始人 & CEO。十年攒下的问题和答案，都汇进了这家公司。我们做在上下文中学习的具身模型，从示范、语言和触觉里学，向人学，也为人服务。
// 03Selected research部分研究
- DualWorld: a dual-system world model for full-size humanoid robots
- Forging productive human-robot partnerships through task training
- Learning a group-aware policy for robot navigation
- Evaluation of socially-aware robot navigation
- PATI: a projection-based augmented table-top interface for robot programming