An analysis argues that three recent moves let foundation models absorb layers world models and integrators once owned: OpenAI's GPT-6 Astra being used by third parties as an action policy on real robots, Google's Gemini 4 Argon generating MuJoCo digital twins and Real2Sim pipelines from video, and Physical Intelligence's Pi API Beta hosting robot fine-tuning and inference. Cited figures include Astra scoring an average success rate in the twenty-percent range across more than forty dual-arm tasks in RoboDojo without task-specific fine-tuning, Argon reaching 87% on the ERQA embodied reasoning benchmark, and a downstream VLA's real-robot success rate rising from the mid-fifties to the mid-seventies percent. The piece says contact physics, millisecond force control and physical causality remain out of reach, leaving four Chinese 'thin-shell' groups most exposed: wrapper vendors, rule-based integrators, vertical VLA startups and low-quality annotation or simulation shops.

OpenAI·GPT-6 Astra·Google·+5