Google shows Gemini 4 Argon controlling robots and building sims from video

Google published a developers post on 13 October 2026 on using Gemini 4 Argon across the robotics stack. Argon scores 87% on Embodied Reasoning (ERQA), which Google says beats other frontier models on pure reasoning for physical 3D problems. Demos include Boston Dynamics Spot learning to open a foot-pedal bin after practising in MuJoCo, zero-shot arm control with visual grids, closed-loop steering of the Gemini Omni world model, and Real2Sim twins from robot cameras, egocentric human video, or paper sketches. Another demo claims Argon raised a VLA success rate from 56% to 75% and cut completion time from 35s to 24s by synthesising 4,500+ sim episodes from one photo. Figures are Google's own demos and benchmarks. Access is the same Argon rollout: AI Studio and Gemini API (`gemini-4-argon`) on the paid tier, and Antigravity for Ultra subscribers. On Hacker News (about 184 points, 45 comments), the largest thread asks how much more efficient models and chips must get before Argon-class reasoning is practical on a robot's battery; replies split between "years away" and "run it over the network from a nearby data centre". A smaller group are hobbyists building their own robots, including a snow-sweeping prototype and a Unitree humanoid controlled by an LLM, who are encouraged by zero-shot control. One sceptic says it falls short of Figure 3 and Unitree's agile robots. No consensus. Sorted take: impressive demos for robotics teams already on Google's stack; not a plug-and-play product for most operators yet.