Google introduced Gemini 4 Argon for robot control and simulation creation
Gemini 4 Argon combines robot control, simulation creation and training data assessment. According to the article, it is available through paid developer programs from Google. The demonstrations still require detailed human instructions.
Google introduced Gemini 4 Argon, a model intended to control robots, create virtual environments for their training and assess the quality of training videos. According to the article, it is available through paid developer programs from Google.
In one demonstration, the model created a simulation from a photograph of a table, proposed 77 tasks and generated more than 4 500 training episodes for a control model. According to Google, subsequent testing on a real robot showed an increase in the success rate from 56 to 75 % and a reduction in execution time from 35 to 24 seconds. The demonstrated examples show that simulations can also be prepared from sketches and videos.
Evaluation results depend on the type of test. In ERQA, a test focused on physical and spatial reasoning, the model achieved a score of 87 %, according to Google. In AA-Robotics from Artificial Analysis, which evaluates direct control of a real robotic arm across eight tasks, it finished third with 55 points out of 100; the top model, GPT-6 Astra, scored 69 points.
When assessing training videos, the model flags issues such as delays and failed grasps so that these segments receive less weight during training. However, the demonstrated workflows require detailed human instructions and guidance, including instructions for analyzing errors and verifying corrections.
Why it matters
Robotics developers gain paid access to a model for preparing simulations and reviewing training recordings. This allows them to try tasks in a virtual environment before deploying them on a real robot. The reported improvement comes from a specific demonstration by Google; the differing test results and the need for human guidance are reasons to verify the benefits on your own task.
Release card
Gemini 4 Argon
- Inputs
- The model processes photographs, sketches, videos and text instructions. It produces interactive 3D simulations, control code and assessments of training videos.
- Availability
- The model is available through paid developer programs from Google.
- ERQA 87 % According to Google, the model achieved this success rate in a test of 400 questions focused on physical and spatial reasoning without auxiliary tools.
- AA-Robotics 55 ze 100 bodů According to the article, the model placed third in a test involving direct control of a real robotic arm across eight tasks.
- Robot control and control code generation.
- Creating training simulations from photographs, sketches and videos.
- Identifying delays and failed grasps in training videos.
- The demonstrated workflows require detailed human instructions and guidance.
Google presents the model as a unified tool for robot control, training simulation creation and data assessment. The article compares it with other models in tests of physical and spatial reasoning and direct control of a robotic arm.
The card summarizes information from the article and any dated corrections, with a link to the original source. It is not our assessment of the model. It does not yet have a dedicated editorial profile. Model selection and other announcements →
Two audiences, two different impacts
What this means
For individuals
A robotics developer can use the model to analyze a training video and identify delays or failed grasps that affect the quality of the training data.
For a business
Robotics teams have an additional paid option for preparing simulations and training episodes. Work planning also needs to account for detailed instructions and validation of results on real hardware.
DevelopmentCheck the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.