Google DeepMind unveils Gemini Robotics 2 AI platform for humanoid robots

Google DeepMind announced Gemini Robotics 2, a new generation of AI models that can control the full body of humanoid robots. The platform adds vision‑language‑action capabilities that let robots walk, crouch, balance, manipulate tools and perform complex, multi‑step tasks such as watering plants, sorting objects and assembling items. Three complementary models make up the system: a Vision‑Language‑Action model for direct motor control, Gemini Robotics ER 2 for embodied reasoning and long‑range task planning, and Gemini Robotics On‑Device 2 that runs locally on the robot to reduce latency and improve safety.

Demonstrations used the Apollo 2 humanoid robot, showing it could fetch a watering can, place books on a shelf and tie knots with a five‑finger hand. Reported success rates ranged from roughly 46 % to 92 % across different manipulation categories. The launch is presented as a step toward physical artificial general intelligence and includes safety mechanisms for risk detection. The technology is being previewed for early partners through Google AI Studio and the Gemini Enterprise Agent Platform.