Google Releases Gemini 4 Argon for Robotics Control, Simulation, and Policy Improvement

What happened

Google introduced Gemini 4 Argon, a new AI model aimed at physical reasoning tasks for robotics. The model is designed to handle robot control, simulation scene reconstruction, video labeling, and iterative policy improvement. It demonstrated its abilities by guiding a Boston Dynamics Spot robot through tasks after training in a simulation environment and reconstructing scenes from video footage. Gemini 4 Argon is now available to users through Google AI Ultra and the Gemini API.

Key details

• Gemini 4 Argon scored 87% on the ERQA benchmark using reasoning alone.
• The model can direct robots like Boston Dynamics Spot and simpler robotic arms using visual and motion data.
• It enables reconstruction of MuJoCo simulation scenes from teleoperation video.
• A reported use case showed policy fine-tuning success rising from 56% to 75% on a real robot setup after domain-randomized simulation.
• Access is provided via Google AI Ultra subscription and the Gemini API under model id gemini-4-argon.

Why it matters

Robotics developers and researchers now have access to a model that can handle a range of physical reasoning and control tasks, potentially streamlining workflows that previously required multiple systems. This could reduce the time and cost needed for developing, testing, and improving robot behaviors, especially for those using simulation-to-real transfer. Users of Google AI Ultra and Gemini API can immediately integrate these capabilities into their robotics projects, gaining higher benchmark performance and new simulation tools.