Google DeepMind is applying Gemini 4 Argon across the robotics lifecycle, from understanding a scene to practicing in simulation and controlling hardware. The company reports an 87% score on ERQA, a benchmark for physical and spatial reasoning in three-dimensional environments.
In one demonstration, Argon created a MuJoCo digital twin of a foot-pedal bin so a Boston Dynamics Spot robot could practice balancing, opening the bin and dropping in cans before acting on the real object. Another test used repeated camera observations to guide a robot arm as it placed a fallen mug upright and moved a napkin into it.
The model can also steer a video world model through an iterative visual loop. Instead of issuing one fixed prompt, it examines intermediate frames and changes the next action. DeepMind presents that feedback cycle as useful for planning dexterous tasks and improving actions after failures.
These examples are demonstrations from DeepMind, not evidence that Argon can safely control arbitrary robots without engineering safeguards. The practical advance is a common reasoning layer that can connect simulation, visual checks and action planning, while real deployments still need calibrated tools, constrained controls and verification.