Google DeepMind has shown how the Gemini 4 Argon neural network can train the robot with a new physical skill on one video. During the demonstration, the humanoid Apptronik Apollo mastered the game in the Cornhol and, according to the developers, performed seven successful throws in a row.
Cornhol is a game in which participants throw small sacs on a sloping board with a hole. In the experiment, Gemini 4 Argon received a video recording of a person performing throws, analyzed the location of the objects and calculated the movements necessary for the repetition of the action. The neural network then wrote the program code of physical simulation, in which the robot could learn a new skill.
The main feature of the development is the transition from a regular video to a working three-dimensional model of the surrounding space. Gemini estimates the distances, position of the objects, the characteristics of their interaction and physical limitations, after which it creates a virtual environment for training. The robot does not need to mechanically copy human movements. The system should select movements taking into account the design of the device itself, its joints and the capabilities of the drives.
The second experiment was closer to everyday tasks. The four-legged robot Spot from Boston Dynamics learned to throw banks in the trash with a foot pedal. In the Gemini 4 demonstration, Argon first created a digital model of the basket and its mechanism in the physical simulator MuJoCo. Spot worked out the balance on three legs while the fourth pressed the pedal opening the lid. After virtual training, the robot repeated the sequence with the real basket, using a manipulator to move the bank.
This approach allows you to check the movements without the risk of damaging expensive equipment. In a virtual environment, the robot can make mistakes, change trajectories, and practice interaction with objects without expending the resources of mechanical nodes. Particularly difficult is the transfer of skills to reality, since the simulator does not always accurately reproduce friction, mass of objects, the work of sensors and other physical features of the surrounding world.
Google conducted a separate test, starting with one photo of the desktop. Gemini 4 Argon built a three-dimensional simulation for robotic manipulators, formed a program of 77 tasks and created more than 4500 virtual episodes of learning. After using the obtained data, the share of successful actions of the real robot, according to the results of internal tests, increased from 56% to 75%, and the average time of the task decreased from 35 to 24 seconds.
The developers also tested the model's ability to reason about physical space. In the ERQA test, which evaluates the understanding of three-dimensional scenes, object location and physical interactions, Gemini 4 Argon scored 87%. Google claims that the result exceeds the performance of other advanced models. However, the high score in such testing does not prove in itself that the robot will be able to perform equally successfully unfamiliar actions outside the prepared experimental environment.
Training robots in computer simulators has been around for a long time. The developers use virtual polygons to train walking, balance retention and complex manipulation, and then try to transfer skills to real mechanisms. The difference between the new approach of Google is to automate the preparation of the environment itself. Instead of manual simulation of objects, setting up physical parameters and writing a large part of the control code, developers can entrust such a work to a universal language model.
Such automation potentially reduces the time of preparation of robots for new tasks, especially when conditions are constantly changing. For warehouse complexes, production lines and service robotics, the ability to build a training environment by video recording or photography could reduce the handmade volume of engineers. However, Google has not yet provided comparative data on the cost and speed of such training regarding traditional methods.
All published results relate to Google's demonstrations and internal tests. There is no independent evidence of the success of the experiments. The company also failed to disclose full statistics of attempts leading up to the seven successful Apollo throws, nor did it show how sustainably formed skills work when changing conditions. In addition, one video served as the source material for a specific task, rather than replacing pre-training of Gemini itself and the software infrastructure of robots.
Google introduced Gemini 4 Argon on October 13, 2026. In addition to robotics, the new model is focused on programming, complex multi-stage tasks and research. However, it is experiments with physical devices that show one of the most interesting directions of development of universal neural networks. Artificial intelligence is gradually moving from image analysis and program generation to creating virtual physical environments in which machines can prepare for action in the real world. So far, we are talking about controlled experiments, and not about robots that can independently master any work after watching a short video.
Cornhol is a game in which participants throw small sacs on a sloping board with a hole. In the experiment, Gemini 4 Argon received a video recording of a person performing throws, analyzed the location of the objects and calculated the movements necessary for the repetition of the action. The neural network then wrote the program code of physical simulation, in which the robot could learn a new skill.
The main feature of the development is the transition from a regular video to a working three-dimensional model of the surrounding space. Gemini estimates the distances, position of the objects, the characteristics of their interaction and physical limitations, after which it creates a virtual environment for training. The robot does not need to mechanically copy human movements. The system should select movements taking into account the design of the device itself, its joints and the capabilities of the drives.
The second experiment was closer to everyday tasks. The four-legged robot Spot from Boston Dynamics learned to throw banks in the trash with a foot pedal. In the Gemini 4 demonstration, Argon first created a digital model of the basket and its mechanism in the physical simulator MuJoCo. Spot worked out the balance on three legs while the fourth pressed the pedal opening the lid. After virtual training, the robot repeated the sequence with the real basket, using a manipulator to move the bank.
This approach allows you to check the movements without the risk of damaging expensive equipment. In a virtual environment, the robot can make mistakes, change trajectories, and practice interaction with objects without expending the resources of mechanical nodes. Particularly difficult is the transfer of skills to reality, since the simulator does not always accurately reproduce friction, mass of objects, the work of sensors and other physical features of the surrounding world.
Google conducted a separate test, starting with one photo of the desktop. Gemini 4 Argon built a three-dimensional simulation for robotic manipulators, formed a program of 77 tasks and created more than 4500 virtual episodes of learning. After using the obtained data, the share of successful actions of the real robot, according to the results of internal tests, increased from 56% to 75%, and the average time of the task decreased from 35 to 24 seconds.
The developers also tested the model's ability to reason about physical space. In the ERQA test, which evaluates the understanding of three-dimensional scenes, object location and physical interactions, Gemini 4 Argon scored 87%. Google claims that the result exceeds the performance of other advanced models. However, the high score in such testing does not prove in itself that the robot will be able to perform equally successfully unfamiliar actions outside the prepared experimental environment.
Training robots in computer simulators has been around for a long time. The developers use virtual polygons to train walking, balance retention and complex manipulation, and then try to transfer skills to real mechanisms. The difference between the new approach of Google is to automate the preparation of the environment itself. Instead of manual simulation of objects, setting up physical parameters and writing a large part of the control code, developers can entrust such a work to a universal language model.
Such automation potentially reduces the time of preparation of robots for new tasks, especially when conditions are constantly changing. For warehouse complexes, production lines and service robotics, the ability to build a training environment by video recording or photography could reduce the handmade volume of engineers. However, Google has not yet provided comparative data on the cost and speed of such training regarding traditional methods.
All published results relate to Google's demonstrations and internal tests. There is no independent evidence of the success of the experiments. The company also failed to disclose full statistics of attempts leading up to the seven successful Apollo throws, nor did it show how sustainably formed skills work when changing conditions. In addition, one video served as the source material for a specific task, rather than replacing pre-training of Gemini itself and the software infrastructure of robots.
Google introduced Gemini 4 Argon on October 13, 2026. In addition to robotics, the new model is focused on programming, complex multi-stage tasks and research. However, it is experiments with physical devices that show one of the most interesting directions of development of universal neural networks. Artificial intelligence is gradually moving from image analysis and program generation to creating virtual physical environments in which machines can prepare for action in the real world. So far, we are talking about controlled experiments, and not about robots that can independently master any work after watching a short video.
