166: 许华哲再次具身创业：不想错过最大的西瓜

Summary

This LateTalk episode interviews Xu Huazhe about leaving Xinghaitu and founding Poke Robotics in March 2026 to pursue general household robots. Xu frames the opportunity as Physical AGI rather than a humanoid shell, an industrial arm, or a narrow scene robot: the valuable layer is a general robot brain that can act across messy physical environments. The source extends the wiki’s Embodied AI branch by adding AI Native Robotics, Unified Robot Models, and Robot Active Use Metrics as route-level claims for household robotics.

Key Claims

• Xu Huazhe says Physical AGI is the real target: a robot should eventually complete many different tasks across scenes, not only optimize one industrial workflow.
• He expects meaningful household robot entry within 18 to 24 months from the May 2026 interview, while still treating full generality as a longer path.
• The episode creates a route tension with industrial-first embodied-AI companies: Xu argues that early logistics, manufacturing, or “robot as equipment” paths can pull attention away from the larger intelligence prize.
• Poke Robotics will start with home robots because the home forces generalization, active agency, task stitching, safety design, and product definition in a way closed industrial demos do not.
• Xu argues that AI Native Robotics is not traditional robotics, not self-driving-style small-scene closure, and not many small deep-learning models stitched together.
• He expects video data, reinforcement learning, world models, and robot data to matter together, but says data quality, failure data, and suboptimal data should be selected and used deliberately rather than mixed blindly.
• Unified Robot Models are central to the route: solving one hundred tasks one by one does not automatically teach the robot how to solve the next thousand.
• The first product stage is to build a hardware body, train AI models, and define the product together; the likely near-term body is a wheeled dual-arm humanoid-like form rather than a fully bipedal robot.
• Safety will be handled partly through product boundaries: early robots should avoid high-risk direct body-contact services such as feeding, lifting infants, massage, or turning people over.
• Xu says shipping units is a weak proxy for real value; Robot Active Use Metrics such as active rate and repeated daily use should matter more than sales volume, dance demos, or short-term production counts.
• Selling robot data to overseas competitors is treated as strategically dangerous because high-quality physical data may be the scarce asset that defines the future robot brain.
• Xu presents “scientist entrepreneurship” as an AI-era advantage when the company depends on technical conviction, taste, and willingness to pursue a non-consensus route.

Key Quotes

“最大的西瓜” — Xu’s phrase for the intelligent robot brain layer he does not want Chinese teams to miss.

“时代在召唤” — Xu’s explanation for why he chose to start again despite family and career pressure.

“不是 Robotics” — his shorthand for why AI-native robotics should not simply inherit traditional robotics task decomposition.

Connections

• LateTalk, Xu Huazhe, and Poke Robotics — show, guest, and new company.
• Physical AGI, Physical AI, and Embodied AI — general physical intelligence framing.
• Home Service Robots, Consumer Robotics Full Stack, and Household Robot Data Flywheel — household product and data loop.
• AI Native Robotics, Unified Robot Models, Real Robot Data Strategy, Vision Language Action Models, World Models, and World Model VLA Fusion — model and data route.
• Robot Active Use Metrics, Product Led Willingness To Pay, AI Consumer Growth Metrics, and AI Commercialization Pressure — commercialization and usage discipline.
• Xinghaitu, Physical Intelligence, Generalist, and Figure AI — comparison companies or adjacent embodied-AI routes named in the episode.
• Research Taste and AI Organization Design — scientist-founder and team-building implications.

Contradictions

• No direct contradiction found. The episode extends prior Xu Huazhe material by moving his departure from Xinghaitu from a planned 2C direction into a concrete Poke Robotics thesis.
• Productive tension to track: Xinghaitu and Production Robot Scenario Selection emphasize production scenes and whole-machine value-chain execution, while Xu argues that over-indexing on early industrial landings risks missing Physical AGI.
• Open issue: the 18-to-24-month household-robot timeline is a forecast from May 25, 2026; the wiki should revisit it around late 2027 to early 2028.