On September 25, 2025, Google DeepMind released Gemini Robotics-ER 1.5, a model designed for use in robotics. This is the first version of the reasoning system publicly available to developers .
The model functions as a ” high-level brain” for the robot : it understands natural language commands, plans multi-step actions, and organizes complex behavioral chains .
It combines visuospatial perception, progress assessment, and the ability to access third-party tools, including Google Search, the Robot API, or VLA (vision-language-action) models.
You can start working with the system right away via Google AI Studio and the Gemini API . A technical report with detailed test results has also been published on the blog.
One of the key innovations is the ability to perform fast and accurate spatial reasoning with low latency. Gemini Robotics-ER 1.5 can determine the coordinates of objects based on their size, weight, and properties and use this data for precise manipulation.
The model can recognize complex tasks such as ” cleaning up a desktop using the example in the photo” or ” sorting waste according to local regulations,” while also consulting online sources for additional information. To execute these scenarios, spatial-temporal planning and success verification are used.
The developers highlight a feature called ” thinking budget ” that allows them to manage the balance between response speed and reasoning accuracy. For simple operations, they can get a quick answer, while for multi-step problems, the model will take longer to “think.”
Great attention has been paid to safety. Gemini Robotics-ER 1.5 better recognizes activities that violate the robot’s physical limits, such as exceeding its payload capacity. Filters are also integrated to prevent the generation of dangerous or malicious plans . The evaluation was conducted, among other things, using the ASIMOV benchmark .
Developers emphasize that built-in safety features must be complemented by traditional engineering methods: emergency stop systems, collision avoidance, and risk analysis.
In practice, the model demonstrates an understanding of action sequences. In experiments, it described the operation of two robotic arms that transfer objects into containers, with precise timing.
In other examples, Gemini Robotics-ER 1.5 planned the coffee-making process : determining where to place the cup and capsule and plotting the trajectory for closing the coffee machine’s lid.
Developers can control the depth of reasoning via API parameters, use ready-made Colab notebooks, and connect the Python SDK for integration into their own projects.
Gemini Robotics-ER 1.5 is positioned as a fundamental component for the creation of future robotic systems in which artificial intelligence will combine perception, planning and physical action.
Follow us on Google News to receive daily updates on cybersecurity. Contact us if you would like to report news, insights or content for publication.
