Google's Gemini 4 Argon takes on the full robotics pipeline

Google had released Gemini 4 Argon globally just days before this robotics-focused pitch, and the company frames it as another step toward "physical AGI," a goal it and its AI subsidiary DeepMind have pursued for years. The demos lean on simulations to prepare, verify or generate training data — a pattern Google applies from a single photo of a table to a first-person video of a human grasping task. The model also works from other sources, such as a teleoperated robot recording, a ballpoint pen sketch, or even directing an AI video world model called Gemini Omni step by step. In one demo, Argon had to place a Lego brick into a wooden box and improve its own performance over time, cutting execution time from 58.9 to 17.4 seconds across three iterations.

Google is positioning Argon alongside its existing robotics work. It had already introduced Gemini Robotics-ER 1.6 for spatial perception and task planning, and training in virtual environments is not new — the startup World Labs showed how real scenes can be turned into virtual training worlds, while Nvidia's DreamDojo simulates robot actions as videos. Google's stated differentiator is the ambition to handle simulation, control and evaluation all with the same model.

That ambition comes with visible caveats. The published example prompts reveal that the demos still need a lot of human guidance — for righting an overturned cup described as "zero-shot," Argon receives a camera image with a centimeter grid overlay and pre-built commands like "grasp" or "lower until contact." For the simulations, there are long prohibition lists to keep the model from taking shortcuts like copying measured sensor values, and Google doesn't share failure rates for individual demos or the total number of attempts behind the 75 percent success rate.