Gemini 4 Argon Takes Apollo From Human Video to Seven Straight Cornhole Shots

Key Takeaways

Show all

• Google says Apollo sank seven consecutive cornhole shots after Argon used human video, simulation practice and real-world feedback to develop the throw.
• Argon’s reported 87% ERQA score measures embodied reasoning questions, separately from physical robot performance.
• Google’s examples span real hardware, physics simulation and generated video; each demonstrates a different part of the development process.

Google DeepMind is using Gemini 4 Argon to connect video understanding, simulation building and physical robot control. In an internal demonstration announced on October 13, Google says Apptronik’s Apollo 2 humanoid learned a cornhole throw through a video-to-simulation workflow and then sank seven consecutive shots on real hardware.

The demonstration accompanies Argon’s wider release and a robotics developer post covering control, simulation reconstruction, data annotation and iterative improvement.

Watching, practicing and adjusting

In the accompanying demo, researchers describe showing Gemini a human cornhole video and prompting it to practice in MuJoCo. The model reviews unsuccessful attempts and revises its approach.

A segment presented as the model’s own summary describes mapping the human arm swing and release onto the robot, locating the board using its cameras, and building a simulation. It then widens the stance and adjusts the swing to clear the hip. On hardware, the account describes further release adjustments based on where bags land.

That last step matters: the demonstrated workflow includes feedback from physical attempts. The seven-shot streak does not establish a first-attempt success rate. Google has not provided a total attempt count or a breakdown of the human engineering behind this cornhole experiment in the linked posts.

Google says the pipeline can reduce weeks of manual simulation modeling to hours. The practical promise is a shorter cycle between observing a task, constructing a test environment and improving the resulting movement.

Reasoning beyond task planning

Russell Mendonca, a Google DeepMind robotics researcher whose profile lists previous work on Tesla’s Optimus AI team, said Argon had changed his expectations for language models in robotics.

“Reasoning can be a powerful mechanism for generalization,” he wrote, extending the point to low-level control as well as semantic planning.

The published examples still involve robotics infrastructure. Google’s arm-control demonstration, for example, gives Argon a calibrated visual grid and predefined motion commands. The materials do not establish that the general-purpose model directly replaces every high-frequency controller.

In a second post, Mendonca described an iterative planning loop using the Gemini Omni world model. Argon writes a plan, inspects the generated outcome and revises it. Google illustrates this with a video rollout of hands arranging tiles to spell ARGON. That is a generated video plan, distinct from Apollo’s physical trial.

A broader development toolkit

Other examples include Spot practicing foot-pedal bin disposal in MuJoCo and Argon reconstructing simulation environments from camera recordings or sketches. Google also reports generating more than 4,500 simulated episodes for VLA fine-tuning, raising real-world task success from 56% to 75% and reducing completion time from 35 to 24 seconds in that evaluation.

The reported 87% on ERQA measures something different. The benchmark repository describes 400 multiple-choice questions combining images and text about spatial reasoning and real-world scenarios. It is a reasoning test, not an 87% robot task-completion rate.

We previously covered Gemini Robotics 2’s whole-body control and multi-robot coordination. Argon extends the picture with examples of a general-purpose model building simulations, generating control code and refining actions through feedback.

The announcement follows DeepMind’s plans for a new Cambridge-area robotics lab. Argon adds a concrete software direction to that expansion: using a general-purpose model to help construct, test and refine the systems that make robots act.