沈宇军 / Shen Yujun

Overview

Shen Yujun / 沈宇军 is the 蚂蚁灵波 / Ant Lingbo chief scientist in the wiki’s embodied-AI branch. His core position is that robot models should be native to physical sensors, spatial perception, action, and real-time execution rather than borrowed wholesale from digital-world language or video models.

Current Profile

Shen’s earlier interview defines the robot-brain-first route; the later Shizilukou Crossing debate adds sharper qualifications about data. He treats pretraining, real-world robot collection, and simulation as stage-dependent, while stressing that real sensors and physical imperfection are difficult to simulate away. He also argues that robots differ from chat systems because they must act while continuously receiving input.

Key Characteristics

• Frames Robot Data Scale Up as the missing condition for a robot GPT-1 moment.
• Treats real-machine data as central, while allowing eGo-style human data and simulation to contribute at different stages.
• Emphasizes sensor noise, tactile data, physical imperfection, and continuous input as robot-native problems.
• Expects robot brains, bodies, sensors, tactile sensing, dexterous hands, and data systems to rise in alternating waves.
• Distinguishes intent execution from autonomous intent origin.

Evidence

• Career evidence: 147. 和蚂蚁灵波沈宇军聊：机器人原生基础模型、大脑和本体的关系、预训练与数据scale up、老师汤晓鸥 presents his path from Tsinghua University / 清华大学, SenseTime, 香港中文大学 / Chinese University of Hong Kong, 汤晓鸥 / Tang Xiao’ou, and ByteDance into robotics.
• Model evidence: 147. 和蚂蚁灵波沈宇军聊：机器人原生基础模型、大脑和本体的关系、预训练与数据scale up、老师汤晓鸥 attributes Embodied Native Foundation Models and the robot GPT-1 data-scale gap to Shen.
• Data-stage evidence: 当具身智能走到十字路口｜对谈苏度、蚂蚁灵波、自变量、破壳：四种一线判断 records Shen saying different data types work at different stages, with internet and real physical data important for pretraining and simulation useful for specific tasks.
• Sensor evidence: 当具身智能走到十字路口｜对谈苏度、蚂蚁灵波、自变量、破壳：四种一线判断 records his tactile-sensor example about simulation mismatch in frequency, amplitude, and consistency.
• Continuous-input evidence: 当具身智能走到十字路口｜对谈苏度、蚂蚁灵波、自变量、破壳：四种一线判断 records his view that robots must update strategy while acting, unlike turn-based digital chat systems.

Qualifications

Shen’s real-data emphasis remains in tension with simulation-heavy and structured-data routes. The source does not settle how much simulation can close the gap as simulators improve.

What Changed

• Converted the page to the synthesis-v1 entity schema.
• Added the four-company debate as evidence for Shen’s data-stage, sensor-noise, and continuous-input claims.

Relationships

• 蚂蚁灵波 / Ant Lingbo - company where he leads the robot-brain effort.
• 蚂蚁集团 / Ant Group - parent-company context for the embodied-AI bet.
• 汤晓鸥 / Tang Xiao’ou, SenseTime, Tsinghua University / 清华大学, 香港中文大学 / Chinese University of Hong Kong, and ByteDance - career and research context.
• Embodied Native Foundation Models - model route he articulates.
• Embodied Robot Data Tradeoff - data debate where he supplies the real-sensor position.