Gemini 4 Argon Writes Physics Simulation Code From Human Video, Teaches Apollo Humanoid Cornhole

Last week, Google DeepMind showed that its newest advanced AI model can watch a video of a person throwing bean bags and use it to teach a robot to do the same thing. The AI figures out the shape and layout of the area, works out how a six-foot robot needs to move its joints, and writes the computer code needed to create a realistic physics simulation for training. It can do all this using a normal camera video, without special motion-tracking equipment or weeks of work by engineers. The result was Apptronik's Apollo humanoid robot making seven cornhole shots in a row. If this method works for many other tasks, it could mean something much bigger. Google already owns the physics simulation software most often cited in robotics research. Now it also has an AI model that can automatically write code for that software using ordinary videos. This could give Google control over many of the important tools that robot makers around the world rely on.

Gemini 4 Argon was announced on September 30, 2026, and became widely available on October 13. It is the first model in Google's Gemini 4 family and the most powerful model Google has made publicly available so far. Argon was not built just for robots. It was designed to handle difficult problems that need many steps of thinking, including software development, legal work, cybersecurity, and scientific research. What makes the cornhole demonstration special is that Google did not need to give Argon extra training specifically for robotics. Instead, Argon used its normal abilities to understand images and videos and write code. With just one request to the model, it used an ordinary video to help teach a real robot a new movement skill.

What Is Gemini 4 Argon?

Argon has an output context window of one million tokens, much larger than the 64,000-token limit of earlier Gemini models. Tokens are small pieces of text that AI models use to process and produce information. Argon is available through Google AI Pro for $20 a month, Google AI Ultra for $99 a month, and the Gemini API for business developers. The API currently costs $2 for every million input tokens and $10 for every million output tokens at introductory prices. These prices are scheduled to double on December 1, 2026, to $4 and $20, respectively.

Google plans to use Argon as one part of a larger robot AI system. Argon does the difficult thinking work in the cloud, such as studying videos, understanding the positions of objects, and writing physics simulation code. Another AI model, called Gemini Robotics ER 2, runs directly on the robot and controls movements that must happen quickly. This setup means Argon does not have to control every robot movement instantly. A large AI model running in the cloud would not be fast enough to do that.

Video-to-Sim-to-Real: Four Steps From Video to Robot Training

Google DeepMind calls its method "Video-to-Sim-to-Real." In the past, engineers often spent weeks or months building physics simulations, writing special code for each robot task, and collecting large amounts of training data from robots. Google's new method combines these jobs into a process that the company says can take only hours.

The process has four stages. First, Argon watches a normal camera video of a person doing the task that the robot needs to learn. It does not need special sensors or equipment that records human body movements. Second, Argon studies the space and the movements. It works out the shape of the scene, finds objects and their positions in three-dimensional space, and decides how the robot's joints should move to copy the action. An important part of this step is changing the movement to fit the robot's body instead of a human body. Apollo has a humanoid arm system that can move in 22 different ways and operates at 500 Hz, or 500 times per second. Argon adjusts things such as how far apart the robot's feet stand and how its arm swings. This helps stop the robot's arms from hitting its hips, a problem caused by the robot's body shape that a human doing the same movement would not have.

The third stage is what makes this system especially different from earlier methods. Argon writes the code for a physics simulation in MuJoCo by itself. MuJoCo stands for Multi-Joint dynamics with Contact. It is free, open-source software that simulates how objects and moving parts behave in the physical world. Emanuel Todorov, Tom Erez, and Yuval Tassa created it at the University of Washington in 2012. Google DeepMind bought MuJoCo in October 2021 and released it as open-source software under the Apache 2.0 license in May 2022. It is the most widely cited simulation tool in robotics research. Argon writes the full XML description of the simulated scene. It also sets the masses of objects, how their weight affects their movement, how much friction they have, and how much force the robot's fingertips should use. Before this, an experienced simulation engineer usually had to set up these details using carefully written instructions.

Google DeepMind also owns mujoco_warp, a version of MuJoCo that uses GPUs to run large numbers of simulations faster. Google therefore owns both a widely used robotics simulation tool and, according to the article's claim, the only AI model that can automatically create code for that tool from ordinary video. This gives Google control over several connected parts of the process. A competitor could not necessarily copy Google's complete system just by building an AI model that gets similar test scores. It would also need to copy or replace the simulation technology that supports the model.

The fourth stage is called system identification, or sysID, followed by transferring what the robot learned in simulation to the real world. Argon repeatedly adjusts the simulation to make it behave more like the real robot. It compares things such as fingertip forces and joint positions in the simulation with measurements recorded by sensors on the physical robot. This helps make the virtual robot match the real robot before the learned movements are used on actual hardware. Norman Di Palo, a Google DeepMind researcher, explained the process on X. He said the team asked Gemini Argon to teach Apollo to play cornhole, and it succeeded without being given task-specific examples for training beforehand. The team only provided a video of a person playing. Argon then took control of the robot, built its own simulation environment, adjusted the simulation to match the real robot, and trained the robot to throw bean bags.

Google also showed the same approach working on a tabletop task. Argon received just one photograph of a tabletop and created a complete virtual copy of it. It then automatically made more than 4,500 training episodes. Across these training runs, it changed object weights, friction, lighting, and starting positions to help the robot practice under different conditions. After training, the real robot arm's success rate increased from 56 percent to 75 percent. The time needed to finish the task also fell from 35 seconds to 24 seconds. In another demonstration, Argon worked with Spot, the four-legged robot made by Boston Dynamics. Before Spot tried the task in real life, Argon tested whether it could balance on three legs in a simulation. The real robot then attempted a task with several steps: stepping on the pedal of a trash bin, picking up a soda can, and putting the can inside. This required planning both how the robot would move its body and how it would use its arm.

ERQA at 87 Percent: What Google's Test Results Show

Along with the robot demonstrations, Google reported that Argon scored 87 percent on its Embodied Reasoning benchmark, called ERQA. This is a test with 400 questions about understanding physical objects, spaces, and problems in three-dimensional environments. According to Google's own results, Argon scored higher than Claude Opus 5.5, which got 82.3 percent, GPT-6 Astra at 81.0 percent, and Claude Fable 5.1 at 74.0 percent on the same test.

There are important reasons to be careful about these results. Google created and ran the ERQA test itself, and independent researchers have not yet repeated the results. The cornhole demonstration was also an internal test of just one throwing task. It does not prove that Apollo can handle many different jobs or use its hands skillfully in general. Google did not say how far the robot stood from the cornhole board or how many throws it tried before recording the seven successful shots in a row.

These limits are important, but the test still looks at a meaningful problem. ERQA checks whether an AI model understands physical spaces and objects, something that general AI models working with language and images have often found difficult. If a model can score this highly while also writing working physics simulation code, it suggests that AI is getting better at connecting its ability to think through problems with the practical skills needed to train robots. Independent testing is still needed to find out exactly how strong Argon's performance really is.

Apollo, Apptronik, and Google DeepMind: A Partnership Linked to a $5 Billion Company

The Apollo robot used in the cornhole demonstration is made by Apptronik, a humanoid robotics company based in Austin, Texas. Apptronik was founded in 2016 as a company that grew out of the University of Texas. Its technology has roots in NASA's Valkyrie humanoid robot program. Since it was founded, Apptronik has developed more than a dozen robot platforms. Apollo is the company's robot designed for commercial use. It stands five feet eight inches tall, or 1.73 meters, and weighs 160 pounds, or 73 kilograms. Its size is similar to a human's so it can work in factories and warehouses built for people. It is designed to use existing shelves, tools, and workstations without companies having to rebuild their facilities.

Google DeepMind and Apptronik began their formal relationship through a strategic partnership announced in late 2024. Under the agreement, DeepMind's robotics team planned to develop AI models based on Gemini for use in Apollo robots. The relationship grew stronger in February 2026, when Apptronik completed a $520 million extension of its Series A funding round. The investment valued Apptronik at $5 billion. Google was one of the leading investors, along with Mercedes-Benz, B Capital, and John Deere. Apptronik planned to use the money to make more Apollo robots, increase its workforce beyond 300 employees, and bring its robots into more factories and logistics operations.

In June 2026, Apptronik introduced Robot Park, a group of facilities where teams of Apollo 2 robots keep collecting information about how they interact with real objects and environments. That information helps Apptronik and Google DeepMind improve their AI models. Argon's Video-to-Sim-to-Real method offers another way to collect useful training information. Instead of relying mainly on robots practicing in the real world, it uses the huge collection of videos that already show humans doing different tasks.

Carolina Parada, Google's vice president and head of robotics at DeepMind, has said the long-term goal is to create a world with many useful robots doing useful jobs. She has also pointed out that this technology is still mainly research.

How Does Argon's Method Compare With VIDEOMIMIC, GRAIL, and λ₀?

Google is not the only group trying to teach robots using videos of people. Researchers at universities and companies developed many similar ideas during 2025 and 2026. VIDEOMIMIC, presented at CoRL 2025 by researchers at the University of California, Berkeley, uses video from a single camera to rebuild human movements. It then changes those movements to work on a Unitree G1 humanoid robot, using a training method called reinforcement learning in simulation. Another system, PRISM, developed by researchers at UC Berkeley and Stanford, uses AI-generated videos showing different possible outcomes to turn a small number of real demonstrations into many different artificial training examples. GRAIL, introduced in June 2026 by researchers from NVIDIA and UCLA, creates training data for humanoid robots that need to walk and handle objects. It uses knowledge from video AI models and three-dimensional digital objects, without collecting data from physical robots. Alibaba's λ₀ system, published in October 2026, trained a humanoid robot control model using 500 hours of video recorded from a person's point of view.

What makes Google's Argon demonstration different is how much of the work one general-purpose AI model can do on its own. Earlier systems usually needed several separate tools. One tool would rebuild movements from video, another would change those movements to fit a robot, another would build the simulation, and another would train the robot. In Google's demonstrations, Argon used its combined abilities to understand videos, recognize objects and their meaning, understand three-dimensional spaces, and write code. It handled much of this process through one interaction with the model. It also wrote the simulation code itself instead of needing an engineer to write it by hand.

Kanishka Rao, a principal engineer at Google, described this as an important improvement because one AI model could combine an understanding of the physical world with the ability to write code and reason about what it sees.

Can Argon's Video Training Method Work for Industrial Jobs?

The biggest practical question is whether Argon's method can work for many different tasks, not just the ones Google has demonstrated. Throwing a bean bag is a fairly simple and controlled task. The way the bag moves is mostly predictable, the object is easy to identify, and the surroundings do not change. Factory jobs can be much harder. Robots may need to handle soft or bendable materials, work in crowded spaces where things keep moving, or carefully control how much force they use when touching objects. Google's demonstrations have not yet shown whether Argon can handle these harder situations.

Google's bigger idea is that this method could work on a very large scale. There are enormous numbers of videos online showing people doing everyday activities, including cooking, putting things together, working in warehouses, and teaching others how to complete tasks. Google believes these videos could become a huge source of robot training information without needing expensive systems to collect data from real robots. If a powerful AI model can watch these videos, create useful robot simulations, and reliably transfer what the robots learn into the real world, teaching robots new skills could become much cheaper. The main challenge would then be making the AI smart and capable enough, rather than collecting enough training data.

If this idea works in many different situations, beyond carefully controlled demonstrations, it could greatly affect how quickly humanoid robots learn new jobs and begin working in new places. For now, Apollo making seven cornhole shots after learning from a video is an early demonstration that the idea can work. It also supports a much larger possibility: the enormous collection of videos people have already made might one day become one of the most important sources of training data for robots.

Gemini 4 Argon is available through Google AI Pro for $20 per month and Google AI Ultra for $99 per month. Business developers can use it through the Gemini API, with introductory prices of $2 per million input tokens and $10 per million output tokens. These prices are scheduled to double to $4 and $20 per million tokens on December 1, 2026. Apollo is made by Apptronik in Austin, Texas, and is currently aimed at factory and logistics work.

Frequently Asked Questions

How Does Gemini 4 Argon's Video-to-Sim-to-Real Method Work?

Argon starts by watching an ordinary video of a person doing a task. It does not need special sensors or equipment that records body movements. The AI works out the shape of the space and how a robot's joints would need to move to copy the person. It then writes MuJoCo physics simulation code on its own. This code creates a virtual environment and sets details such as object weights, friction, and how objects move when force is applied. The system runs thousands of practice sessions in the simulation, changing conditions between sessions to help the robot learn. It then compares the simulation with sensor recordings from a real robot and adjusts the virtual version to better match the physical machine. Finally, it transfers the learned instructions to the real robot. Google says this process can take hours instead of the weeks of engineering work that were previously needed.

What Is the ERQA Test, and Can Readers Trust Google's Reported 87 Percent Score?

ERQA stands for Embodied Reasoning. It is a test with 400 questions that measures how well an AI model understands physical problems and the positions of things in three-dimensional spaces. Google reported that Argon scored 87 percent on its Embodied Reasoning test. That was higher than Claude Opus 5.5 at 82.3 percent, GPT-6 Astra at 81.0 percent, and Claude Fable 5.1 at 74.0 percent. But Google designed and ran the test itself. Independent researchers have not yet repeated the results. Until another group confirms the scores under carefully controlled conditions, readers should treat them as Google's own unverified claims. ERQA tests how well AI answers questions about three-dimensional spaces and understands physical relationships. It does not test every skill a real robot needs. A high score therefore does not prove that a robot can handle many different physical tasks.

Why Does It Matter That Google Owns MuJoCo, the Simulation Software Argon Uses?

MuJoCo is the most widely cited physics simulation tool in robotics research. Universities and companies around the world use it to test robots in virtual environments. Google DeepMind bought MuJoCo in October 2021. Gemini 4 Argon can now write MuJoCo simulation code automatically after watching ordinary videos. This means Google controls both an important simulation platform used by robotics researchers and, according to the article, the only AI model able to create simulations for it automatically from video. By controlling these connected parts of the technology, Google has an advantage. A competitor would not necessarily be able to copy Google's complete method simply by creating an AI model with similar test results. It would also need to create or replace the simulation technology that makes the process possible.

Can Robots Trained This Way Do Real Factory Jobs, or Does It Only Work for Cornhole?

That question has not been answered yet. Throwing a bean bag is a controlled task where the movements and results are fairly predictable. The tabletop demonstration, where the robot's success rate improved from 56 percent to 75 percent, suggests the method might be useful for industrial work. However, Google has not shared enough public details about that test for others to judge it fully. Real factory jobs often involve bendable materials, crowded workspaces, changing amounts of force when objects touch, and the need to succeed very reliably. The published demonstrations have not tested these challenges. For businesses thinking about using this technology, the most important unanswered question is whether it can work reliably across many different industrial tasks and when it might be ready to use in their own robot systems.