AI Native Robotics

AI native robotics is Xu Huazhe’s route claim in 166: 许华哲再次具身创业：不想错过最大的西瓜 for building robots around general model capability rather than around traditional robotics decomposition. In his framing, “AI native” does not only mean adding neural networks to robots; it means letting model generalization, data diversity, active agency, and product definition shape the entire stack.

Xu defines the idea partly by exclusion. It is not traditional robotics that solves one closed manipulation problem at a time, not autonomous-driving-style closure inside a narrow scene, and not many small deep-learning models patched together into a brittle imitation of general intelligence.

Key Claims

• Data diversity matters because a closed environment can generate many examples without teaching enough general physical structure.
• Household robotics raises the bar because homes contain open object sets, changing layouts, ambiguous instructions, and multi-step task gaps.
• Unified Robot Models are the preferred direction because task-by-task models may not transfer into broader competence.
• Product safety boundaries are part of the AI-native route: the model can improve inside constrained tasks before it is trusted with direct body-care services.

Connections

• Physical AGI — the higher-level goal.
• Poke Robotics and Xu Huazhe — source company and founder.
• Vision Language Action Models, World Models, and World Model VLA Fusion — model families whose relationship remains open.
• Real Robot Data Strategy and Household Robot Data Flywheel — data layer.
• Consumer Robotics Full Stack — hardware and product layer that has to co-evolve with the model.