Google DeepMind has introduced Gemini Robotics 2, its next-generation AI system designed for robotics. The new model enables robots not only to understand natural language instructions but also to perform complex physical tasks, plan multi-step actions, and collaborate with other robots.
The new platform consists of three core models:
- Gemini Robotics 2 – a Vision-Language-Action (VLA) model capable of controlling an entire humanoid robot, from head to toe.
- Gemini Robotics ER 2 – a reasoning agent designed to plan and execute complex, multi-step tasks that can last several minutes.
- Gemini Robotics On-Device 2 – an on-device model that runs locally without cloud connectivity and can adapt to new robot bodies within just a few hours.
Unlike previous generations that primarily focused on robotic arms and upper-body manipulation, Gemini Robotics 2 extends AI control to full-body movement. Robots can now walk, bend, squat, lift objects, and navigate their environments while responding to natural language commands. During demonstrations, Google showcased Apptronik’s Apollo humanoid robot performing tasks autonomously using the new system.
Google also highlighted major improvements in robotic dexterity. The model enables robots equipped with five-fingered hands to perform delicate tasks such as tying knots, sealing packages, and replacing light bulbs. It is also compatible with simpler two-finger industrial grippers, making it suitable for a wide range of robotic platforms.
Meanwhile, Gemini Robotics ER 2 gives robots advanced reasoning capabilities. Instead of reacting to individual commands, the AI can break down complex objectives into smaller steps, adapt its plan when conditions change, and recover from errors without requiring constant human intervention. The system also supports multi-robot collaboration, allowing different machines to coordinate and share workloads.
Google says safety remains a key priority. Gemini Robotics 2 is designed to detect nearby humans, refuse potentially dangerous commands, and request human assistance when it encounters uncertain situations.
Rather than building separate AI systems for every type of robot, Google DeepMind aims to create a universal AI brain that can power humanoid robots, industrial manipulators, and service robots alike. If successful, Gemini Robotics 2 could become a common intelligence layer across a broad range of robotic platforms.
While companies such as OpenAI, Anthropic, and xAI continue to focus primarily on digital AI assistants and software agents, Google DeepMind is making a significant push toward embodied AI. Gemini Robotics 2 represents another major step toward robots that can understand, reason, and act autonomously in the physical world, bringing the industry closer to general-purpose intelligent robots.
