Gemini 4 Argon Helps Robots Learn Cornhole Throws and Trash Disposal From Ordinary Video

Google DeepMind showed this week that Gemini 4 Argon can watch a short video of a person tossing beanbags into a cornhole board, map the arm swing and release point onto a humanoid’s joints, build a full physics simulation of the robot and the board, practice the motion until the hand clears the hip cleanly, then transfer the result to Apptronik’s Apollo humanoid. On real hardware the robot sank seven consecutive shots. Researchers described the sequence as turning ordinary video into a scalable training ground that compresses weeks of manual simulation modeling into a few hours.

Argon scored a remarkable 87% on the Embodied Reasoning benchmark, which included of spatial and physical problems asked in 3D situations. Google’s findings demonstrate that when Argon was tested without any particular robot-specific adjustments, it outperformed every other cutting-edge AI model they tested on the identical set of questions. In addition, it is responsible for managing a Boston Dynamics Spot quadruped with a six-degree of freedom arm. The Spot robot steps on a pedal bin, holds a soda can in its hand, checks its wrist cam to ensure that it is centered, and then drops the can into the bin. Before making a physical effort, Argon constructs a computer clone of the bin in MuJoCo, a simulation environment, and tests it for balance on three legs. Only then does it actually move the pedal.

A single snapshot of a table is all it takes for Argon to create a digital duplicate of the table in MuJoCo, come up with new tasks, and devise a strategy for the robot to perform over 4500 possible variations on the task before even training a new vision-language-action model. With that model in place, the bot improves its success rate on the physical robot from 56% to 75%, and it does so 11 seconds faster. Argon can figure out the simulation and its behaviors using footage from the robot’s camera or even a crude design of a humanoid robot knocking over a box, and the real robot can be involved before it has a chance to muck things up.

Argon can also retro-label films of the robot in motion with impressive precision, even down to how its fingers touch with objects. It can also alter and improve its own performance on a six-degree-of-freedom arm, making pick-and-place operations faster and more accurate with each pass. If you want to see it in action, simply give it a prompt and it will recreate various published robotics control papers in a browser-based MuJoCo environment. Access is available today through the paid Gemini API and for Google AI Ultra subscribers.