BREAKTHROUGHS · OCT 13, 2026
Google's Gemini 4 Argon teaches Apptronik's Apollo to play cornhole
Google says its new model watched a human throw, built a simulation, and coached a humanoid to seven straight shots. It is one task, not a chore robot.
By Riley Hart · 2 sources linked below
Nothing changes for home robot owners today. Argon is a developer model sold through Google's API, and Apollo is not a home product. The idea to watch is robots learning a skill from an ordinary video instead of weeks of engineering, which is what a chore robot would need to pick up new tasks in your house.
Google released Gemini 4 Argon on October 13, 2026, and used part of the launch to show it controlling a humanoid robot. Humanoid.guide reports the model became available to developers and businesses through Google's paid API the same day.
In the demo, researchers gave Argon a video of a person playing cornhole. Google says the model mapped the human arm swing onto Apptronik's Apollo humanoid, located the board with cameras, wrote a physics simulation in MuJoCo, practiced there, and then moved the result to the real robot. Apollo then landed seven shots in a row.
Google calls the workflow Video-to-Sim-to-Real and claims it can cut weeks of hand-built simulation work down to hours. Humanoids Daily notes that Google gave no total attempt count, so the seven-shot streak does not reveal a success rate. The arm control also relied on a calibrated visual grid and preset motion commands.
Google also reported 87% on ERQA, a 400-question spatial reasoning test. Humanoids Daily points out that this is a reasoning score, not a rate of finished robot tasks. In a separate evaluation, Google says adding more than 4,500 simulated practice runs raised real-world task success from 56% to 75%.
Developer pricing starts at $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after November 30, 2026, according to Humanoid.guide. Google has not said when, or whether, any consumer robot will run Argon.
Learning from video matters for homes because every house is different. Home robots on our tracker, such as Flourish 1 and Weave's Isaac 1, still depend on demonstrations or remote helpers for new chores.