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

Summary

This Shizilukou Crossing episode puts four embodied-intelligence operators into one debate: Han Zheng / 韩正 from 速度科技 / Sudu Technology, 沈宇军 / Shen Yujun from 蚂蚁灵波 / Ant Lingbo, 王前 / Wang Qian from 自变量 / Zibianliang, and Xu Huazhe from Poke Robotics. The conversation frames 2026 robotics around three unresolved disputes: how to combine simulation and real robot data, whether general foundation models can subsume embodied companies, and which scenarios prove real commercialization rather than demo heat. Its synthesis is that physical-world robotics needs a layered system of data, models, reliable low-level skills, sensors, hardware, deployment economics, and customer renewal rather than one single winning route.

Key Claims

• Simulation is useful for scale, reinforcement learning, and basic manipulation-skill pretraining, but real robot data still carries sensor noise, contact irregularity, tactile signal variation, and physical imperfections that are hard to reproduce.
• Astra-style general models may be strong at semantic and spatial understanding, but complex contact, continuous sensor feedback, and high-success physical execution remain the defensible problem space for embodied-AI companies.
• Layered Robot Architecture is becoming a clearer shared route: a high-level model can understand scenes and decompose tasks, while a lower-level embodied model must execute generalized short skills reliably.
• Commercial robotics is judged less by impressive demos than by customer repurchase, recurring payment, deployment speed, ROI, high first-pass success, and low post-training cost.
• The guests disagree on which bottleneck is most overestimated: raw data volume, model-training tricks, supply-chain maturity, or the idea that large digital models can naturally solve physical control.
• The long-run case for embodied intelligence is that AI must eventually affect material production, from service work to manufacturing, data-center construction, and complex industrial processes.

Key Quotes

“客户能持续付钱，并且付的钱越来越多” - Wang Qian’s single commercial test.

“复杂接触任务仍是具身机器人公司的最后阵地” - the source’s clearest boundary around general-model entry.

“具身智能是复杂系统工程” - Shen Yujun’s caution against reducing deployment to model training alone.

Connections

• Han Zheng / 韩正, 速度科技 / Sudu Technology, Sim2Real, and Layered Robot Architecture - simulation-led, low-level skill reliability route.
• 沈宇军 / Shen Yujun, 蚂蚁灵波 / Ant Lingbo, Embodied Native Foundation Models, and Robot Data Scale Up - real sensor, continuous-input, robot-native model route.
• 王前 / Wang Qian and 自变量 / Zibianliang - data/infrastructure-centered claim that intelligence lives in data and validation loops.
• Xu Huazhe and Poke Robotics - complex contact, household/service robot, and Physical AGI route.
• Embodied Robot Data Tradeoff, General Model Robot Boundary, and Robot Deployment Success Economics - concepts added by this source.
• Robot Commercialization Negative Feedback Loop, Physical AI Manufacturing Gap, and Physical AI Hard Takeoff - adjacent commercialization, manufacturing, and long-run production frames.

Contradictions

• No settled contradiction found. Productive tension remains between 速度科技 / Sudu Technology’s confidence in scalable simulation, 蚂蚁灵波 / Ant Lingbo’s emphasis on real sensor imperfections, and Poke Robotics’ skepticism toward raw data-volume worship.
• The source also qualifies the wiki’s more aggressive Physical AI Hard Takeoff branch: physical production may be essential to long-run AI power, but near-term deployment still depends on success rate, customer ROI, and system engineering.