Google wants Gemini 4 Argon to be the only AI model robots need
decoder ·
Google DeepMind has introduced Gemini 4 Argon, a unified AI model designed to manage the entire robotics pipeline, including simulation creation, robot control, and training data evaluation. The system can generate virtual practice environments from simple inputs like photos or sketches, allowing robots to refine movements before real-world deployment. In demonstrations, Argon guided a Boston Dynamics Spot robot to toss cans into a bin after practicing in a MuJoCo simulation, significantly improving success rates and execution speed. The model also acts as a director for video world models and evaluates training videos by flagging inefficient sub-steps to optimize future learning. While Argon leads on the ERQA benchmark for spatial reasoning, it ranked third on the AA-Robotics benchmark for direct physical control tasks. Despite these capabilities, the demos reveal that the system still relies heavily on detailed human prompting and specific constraints to function effectively. Google states that Argon represents a step toward "physical AGI" by handling simulation, control, and evaluation within a single framework. The model is currently available to developers through Google's paid subscription tiers.
decoder