Podcast Atlas
Source note Episode guide Original audio Topics: Technology, Economics

于是转身向具身走去｜对话王家伟：24 岁的具身智能首席科学家

Summary

This Shizilukou Crossing episode interviews 王家伟, the 24-year-old chief scientist of 深普智能, following his path from a USTC youth-class program and a joint PhD with Microsoft Research Asia through internships at DeepSeek and ByteDance Seed into embodied-intelligence entrepreneurship. The technical center is that embodied intelligence cannot be one general model: an upper “brain” for reasoning, planning, and world understanding has to be paired with a lower, real-time, controllable Action Policy or “small brain.” The company’s route combines self-built UMI-style data collection, pretraining, a robot Agentic OS built around System 1 and System 2, and its own evaluation instead of a single VLA, world-model, or imitation-learning recipe.

Key Claims

  • A general model such as ChatGPT 6 / Astra can surprise on robot manipulation, but the source uses its sped-up video, slow single-pass inference, and static-scene success to draw the general-model boundary at real-time recovery from physical disturbance rather than at grasping or task planning.
  • 深普智能 defines embodied intelligence through an explicit capability framework: adaptability (zero-shot and few-shot through fully out-of-distribution scenes), steerability (control by language, image, and video), and context understanding.
  • The company positions itself as a full-stack terminal company rather than a data company or a model company, on the argument that value comes from putting data, models, deployment, and the body into one product, and that full-stack work can itself become the moat.
  • Self-built UMI-style collection hardware is treated as a differentiator: a six-view glove with millimeter localization and microsecond synchronization replaces base stations and reduces occlusion, and a trajectory-replay test that found over 95% of collected trajectories reproducible on the body justified scaling collection.
  • Context and memory are treated as first-class capabilities across short, medium, and long horizons, where the short horizon can be a sub-second causal chain and longer horizons hold object, place, and human-partner knowledge; growing token counts then collide with real-time latency.
  • Touch is deliberately staged rather than scaled: the open data carries no tactile signal, vision is judged sufficient for two-finger grip checks, and because tactile sensors differ in resolution and placement, committing to a hundred devices now risks a stranded data format.
  • Robotics evaluation is called out as unreliable because benchmarks skew to simulation, can be gamed by post-training, and are hard to reproduce on real machines; the company says it will publish its own simple/medium/hard evaluation and disclose how much data a zero-shot or few-shot result used.
  • The source reports a scaling signal while refusing a scaling-law claim: data diversity, compute, and volume show power-law-like improvement, but serious verification is missing, and the company’s own cost-controlled UMI data with 90%-plus usability is what makes controlled parameter and data-ratio experiments possible.

Key Quotes

“Keep the world simple” — the slogan the episode associates with 深普智能.

“上层需要具备推理、规划和世界理解能力的大脑，下层还需要可实时反应、稳定可控的 Action Policy，也就是类似’小脑’的能力。” — the source’s central brain/small-brain architecture framing.

“夹松了、杯子下滑、再夹紧” — the source’s example of causal context unfolding inside about one second.

Connections

Contradictions

  • No settled contradiction is recorded. The source is a chief-scientist interview that does not disclose model architecture, training cost, concrete evaluation results, or commercialization details, so its internal-capability claims remain source-scoped.
  • Productive tension worth tracking: this source treats tactile data as a staged option, while Tactile Sensing and Yimu Technology / 一目科技 argue that touch is core infrastructure for dexterous manipulation. The source also qualifies simulation-first and general-model-subsumption positions recorded elsewhere in the wiki without directly refuting them.