DeepMind presented Gemini Robotics: a model designed for real robots with movable arms.
By ForGeeks The company has unveiled two new artificial intelligence models that it says “set the stage for a new generation of useful robots.”
Gemini Robotics is an advanced vision-language-action model (VLA) based on the Gemini 2.0 system. This version includes the addition of physical actions as an output response to a query. For example, on a Pixel phone, Gemini’s “response” would be to perform an action or answer a question. The Gemini in the robot will instead take that command as something it has to physically respond to.
The Gemini in the robot will instead take that command as something it has to physically respond to.
The second model is the Gemini Robots-ER, which is a “vision-language learning” (VLM) with “advanced spatial understanding.” This version of Gemini has “embodied thinking” to help the AI effectively navigate the changing environment. In one of the video examples Google demonstrated for journalists, the robot can distinguish between bowls of different finishes and colors on a table. It is also able to identify artificial fruits, such as grapes and bananas, and distribute them to the appropriate bowls. In another example, the robot understands how to pack granola into a lunchbox, which demonstrates its ability to handle details.
Another example of the robot’s ability to process details.
The main focus of this announcement is Google’s praise of the DeepMind team for creating Gemini as the “brain” for robotics. It further emphasizes the interesting fact that the AI that resides on your smartphone can now control humanoid robots.
Google is partnering with companies like Apptronik to create the next generation of humanoid robots. The Gemini Robots-ER model will also be available for testing from partners including Agile Robots, Boston Dynamics and Enchanted Tools. Although the robots are coming soon, the exact timing of their release is not yet known.
According to the company’s website, the robots will be available to the public soon.
In addition, Google is preparing for questions about safety measures for Gemini. Google explained that Gemini Robotics-ER models are capable of evaluating the safety of performing actions in a given context. This is based on frameworks such as the ASIMOV dataset, which helps researchers assess the consequences of safe robot behavior. Google is also working with safety experts to ensure that AI is developed responsibly.
