tl;dr: Google's Gemini 4 Argon is a multimodal model aimed at robotics, scoring 87% on the Embodied Reasoning (ERQA) benchmark and targeting physical reasoning across the robotics stack. It handles tasks like zero-shot robot control, Real2Sim twin generation from video/sketches, hindsight labeling, dense value function estimation, and self-improvement loops (e.g., boosting a VLA pick-and-place success rate from 56% to 75% via synthesized simulation data). It's available now via Google AI Studio, the Gemini API, and Google Antigravity for AI Ultra subscribers.