Google DeepMind has unveiled a new generation of AI models capable of fully controlling humanoid robots. The Gemini Robotics 2 family includes several sub-models: Gemini Robotics (for direct control), Gemini Robotics ER 2 (embodied reasoning for spatial understanding and task planning), and Gemini Robotics On-Device 2 (runs locally on the robot to minimize network latency). These systems can coordinate complex, whole-body movements — from limbs to fine motor skills like grasping small objects. In demonstrations, robots successfully performed tasks such as placing objects, tying knots, and fastening zippers. The advancements could significantly expand the practical applications of robotics in industries like manufacturing, healthcare, and logistics. Google also introduced a new safety benchmark called ASIMOV-Agentic to evaluate whether models will refuse unsafe tool calls or halt actions when a human is too close.
About the Author
