Google DeepMind presents a new generation of vision-language-action models capable of coordinating walking, posture and object manipulation on a humanoid robot.
Google DeepMind unveiled at the end of July 2026 Gemini Robotics 2, a new version of its vision-language-action model for robots. The company describes the output as a step towards a "physical AI": an intelligence capable of acting in the real world rather than merely producing text or images.
The heart of the announcement holds in what DeepMind calls intelligent control of the whole body. Rather than piloting locomotion and grip separately, the system coordinates all joints to sequence tasks in several steps: walking, squatting, leaning, grabbing objects and storing a crowded room. Several robots can also distribute the work.
The output is structured around three complementary models. Gemini Robotics 2 translates visual and language inputs into motor controls. Gemini Robotics ER 2 ensures incarnate reasoning, task planning and coordination. Gemini Robotics On-Device 2, more compact, runs directly on the robot's on-board hardware without network connection — a determining point for uses where latency or network coverage are problematic.
Read the full article on the original source
Go to the source →