From video to throw: Google's Gemini 4 Argon teaches robots new tasks

Alongside the broad release of its new flagship model Gemini 4 Argon on Tuesday, Google DeepMind published robotics demonstrations. In the best-known example, the model watches a short video of a person tossing beanbags at a cornhole board. Argon maps the arm swing and release point onto the joints of Apptronik's Apollo humanoid robot, builds a physics simulation of the robot and the board and lets it practise there. According to Google, the real robot then sank seven throws in a row.

In a second demo, Argon controls a Boston Dynamics Spot robot with a gripper arm: it steps on the pedal of a bin and drops a drinks can inside. Beforehand, the model checks in the MuJoCo simulation environment whether the robot can keep its balance on three legs. From a single photo of a table, Argon also creates a digital copy with more than 4,500 training variations – which, according to Google, raised the real robot's success rate from 56 to 75 percent.

Google calls the approach "Video-to-Sim-to-Real". On an embodied reasoning benchmark, Argon scores 87 percent according to the company, ahead of other frontier models. Since Tuesday the model has been available through the Gemini API and to Google AI Pro and Ultra subscribers; independent tests of the robotics results are still pending.