Gemini Robotics

History

2025: Product Announcement

In September 2025 Google DeepMind , she introduced the new artificial intelligence Gemini Robotics 1.5 and Gemini Robotics-ER 1.5 models, which train robots to perform complex multi-stage tasks in the real world. The systems allow robots to plan actions before execution, including sorting laundry by color and recycling garbage according to established rules.

Models increase the ability of robots to reason and help solve problems that require several minutes of continuous work. Technology develops the capabilities of robots from executing single commands to a real understanding of physical tasks. The new models follow a series of instructions and use external tools to solve problems. Robots have gained access to Google search for up-to-date information. In demonstration videos, the robot packed a hat and umbrella into a bag for a trip to London after checking the weather forecast. Another example showed the sorting of garbage according to San Francisco rules found on the Internet.

𝓲

Фото: deepmind.google

Senior Director and Head of Robotics at Google DeepMind Carolina Parada celebrated the transition to a new level of understanding. According to her, the models previously coped well with one instruction at a time. Now systems are moving to a real understanding and solution of physical problems in a real environment.

The key innovation was the technology of "movement transfer." The system allows one AI model to use skills developed for a particular type of robot. Robotic arm skills can be applied to a humanoid robot. Previously, robot training was tied to a specific type of device.

The technology significantly expands the amount of training data available. Models can leverage the expertise of different robotic systems to improve performance. The transfer of skills between platforms accelerates the learning process of new robots and reduces development costs.^[1]

Notes