🤖Google Ships Gemini 4 Argon, Scores 87% on ERQA Robotics Benchmark

Your robot sim just got a lot more honest

Google just dropped Gemini 4 Argon, a general-purpose physical-reasoning model built for robot control, simulation, and policy improvement. It's the first Google model we've seen that treats the physical world as the main job, not an add-on to chat. If you're building robot stacks, this lands on your workflow directly. The model can steer a simpler arm using visual coordinates and HTTP motion primitives, and it can close the loop with the Gemini Omni video world model. The Real2Sim examples rebuild MuJoCo scenes from teleoperation video, and labeling prompts annotate demo footage for behavior cloning. The catch is that the scaffolding around the model is part of the system under test, so your prompts, grids, and verifiers matter as much as the weights. The headline numbers: 87% on ERQA using pure reasoning, and a vision-language-action policy that went from 56% to 75% success on the matching real setup after fine-tuning on domain-randomized MuJoCo episodes synthesized from one RGB photo. Access is through Google AI Ultra subscriptions and the paid Gemini API under the ID gemini-4-argon. Google doesn't publish per-demo failure rates beyond that single VLA case, so treat the 75% as a best-case data point until you reproduce it on your own hardware.