Google DeepMind unveils new AI models for controlling robots

Google has launched two artificial intelligence (AI) models, Gemini Robotics and Gemini Robotics-embodied reasoning (ER), built on its Gemini 2.0 foundation, to drive robot capabilities.

In a blog post on Wednesday, Google said Gemini Robotics enhances how robots understand and respond to instructions, enabling them to complete complex, multi-step tasks.

The model integrates vision-language-action (VLA) capabilities, allowing robots to adapt to new environments and handle unfamiliar objects without prior training.

According to the company, Gemini Robotics-ER improves spatial awareness and task planning, helping robots perceive their surroundings, estimate states, and generate code for real-time execution.

Google said the advancement would allow roboticists to incorporate AI into existing low-level control systems.

“Alongside Gemini Robotics, we’re introducing an advanced vision-language model called Gemini Robotics-ER (short for ‘“embodied reasoning”),” the company said.

“This model enhances Gemini’s understanding of the world in ways necessary for robotics, focusing especially on spatial reasoning, and allows roboticists to connect it with their existing low level controllers.”

Google said the models were trained on the ALOHA 2 bi-arm robotic platform but can also operate Franka arms and humanoid robots like Apollo, developed by Apptronik, the robotics company.

The company also announced a partnership with Apptronik to advance humanoid robotics, adding that it is working with other key robotics firms such as Agile Robots, Agility Robotics, Boston Dynamics, and Enchanted Tools to test Gemini Robotics-ER.

Google said the innovations will drive AI-powered automation across industries, improving efficiency, adaptability, and human-robot collaboration.