Google's DeepMind Unveils A New AI Model That Allows Robots To Perform Real-world Tasks Without Training

Mar 29, 2025 Leave a message

Google's DeepMind has launched two new AI models designed to help bots accomplish more real-world tasks.

Gemini Robotics

One of them, called Gemini Robotics, is a visual language action model that enables robots to understand new situations without being specifically trained.

 

Gemini Robotics is based on Google's latest version of its flagship AI model, Gemini 2.0. Carolina Parad, senior director of robotics at Google's DeepMind, has said that Gemini Robotics relies on Gemini's multimodal world understanding capabilities to apply them to the real world by adding new modalities of physical action.

 

The model makes progress in what Google DeepMind believes are the three core areas needed to build efficient bots: versatility, interactivity, and flexibility. In addition to being able to cope with new situations, Gemini Robotics is better at interacting with people and the environment, and is able to perform more precise physical operations, such as folding paper or opening bottle caps.

 

The other is the Gemini Robotics-ER (Representational Reasoning) model, which the company calls an advanced visual language model that is capable of "understanding complex and dynamic worlds."

 

Parada further explains that when you're packing a bento box, you have to consider where and how to do it on the table. Gemini Robotics-ER is designed for this type of inference task, allowing roboticists to interface with existing low-level control systems and unlock new capabilities powered by Gemini Robotics-ER.

 

Vikas Sindhwani, a researcher at Google's DeepMind, said the company was developing a "layered security strategy" and said the Gemini Robotics-ER model had been trained to assess whether an action was safe in a given situation. The company has also released new benchmarks and frameworks to advance security research in the field of AI. According to IT House, last year, Google DeepMind launched the "Robot Constitution", which is a set of robot behavior norms inspired by Isaac Asimov.

 

According to foreign media The Verge, Google's DeepMind has partnered with Apptronik to "build the next generation of humanoid robots". In addition, Google has opened up the Gemini Robotics-ER model to "trusted testers" including Agile Robots, Agility Robotics, Boston Dynamics, and Enchanted Tools. "We're focused on building intelligence that understands and acts in the physical world, and we're very much looking forward to applying this technology across multiple domains and multiple manifestations," Parada said. "