Podcast Atlas
source Episode summary Updated 2026-08-07 Tags: Podcast, Robotics, Embodied-Ai, Physical-Ai, Foundation-Models

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

Summary

This 张小珺Jùn｜商业访谈录 episode interviews [[ShenYujun|沈宇军]], chief scientist of [[AntLingbo|蚂蚁灵波]], about building a robot “brain” before committing to a single robot body. The interview argues that Embodied Native Foundation Models should start from real sensors, spatial perception, video time series, action, and robot data rather than simply adapting digital-world language or video models. Its central bottleneck is Robot Data Scale Up: Shen says embodied AI has not reached a robot GPT-1 moment because the field still lacks a scalable, useful data path comparable to internet-scale language data.

Key Claims

  • [[AntLingbo|蚂蚁灵波]] was set up inside [[AntGroup|蚂蚁集团]] to explore physical-world AI and embodied intelligence, with the strategic choice to build a cross-embodiment robot brain rather than a single proprietary body.
  • [[ShenYujun|沈宇军]] moved from [[TsinghuaUniversity|清华大学]], [[SenseTime|商汤]], [[ChineseUniversityOfHongKong|香港中文大学]], [[TangXiaoou|汤晓鸥]]’s team, image generation, and [[ByteDance|字节跳动]] visual applications toward robotics because robots make vision valuable inside the physical world.
  • Shen argues that a robot-native model stack needs physical-world pretraining: real sensor noise, depth, contactability, spatial geometry, video sequence, action, and real-time execution all differ from clean academic data or digital video generation.
  • The first generation of Lingbo’s brain used depth, video action, world model, and [[VisionLanguageActionModels|VLA]] directions across nine configurations for simpler grasping and desktop tasks.
  • The second generation expands from about 20,000 hours to about 60,000 hours of data, from nine to more than twenty configurations, and from arms/grippers toward head, waist, mobile chassis, wrist cameras, and dexterous hands.
  • The source presents data cleaning as part of scaling: first-generation data had to be reprocessed through a stricter pipeline before the second-generation training set became meaningfully larger.
  • Shen treats real-machine data as the main training source, eGo-style human first-person data as useful, and simulation as stronger for evaluation and partial scenes than for general manipulation training.
  • Robot Data Scale Up depends on task design, body coverage, camera positions, degrees of freedom, and cross-body cleaning, not only on recording more hours.
  • Lingbo’s updated VLA work is valuable partly because real deployment shows what data industrial tasks actually require.
  • The source separates “intent execution” from “intent origin”: Lingbo currently aims to do the instructed task well, not to make robots autonomously infer higher-order goals such as noticing rain and deciding to fetch an umbrella.
  • Shen says the second generation improves random-position and disturbance handling, including billiards and desk-organizing examples, but still lacks strong new-task generalization.
  • He thinks the robot GPT-1 moment has not arrived because embodied data still has no clearly scalable collection route; when embodied data can scale toward internet-data levels, the field will be closer.
  • Shen expects robot brains, bodies, sensors, dexterous hands, tactile sensors, and data systems to rise in alternating waves rather than one layer solving the whole problem first.
  • The episode’s final bet is “embodied-native”: embodied intelligence should have its own models, while large language models remain useful as an instruction-understanding entrance.

Key Quotes

“具身原生” — Shen’s final description of the core bet.

“数据不够” — his diagnosis of the main robot-intelligence bottleneck.

“J 人” — the practical personality he says he would prefer in a robot.

Connections

Contradictions