Google has introduced Gemini Robotics 2, a suite of AI models designed to give robots intelligent whole-body control and the ability to adapt to new tasks and environments. The system represents a significant advance over previous robotic AI, which typically required pre-programming or teleoperation for narrow, repetitive sequences.
The Gemini Robotics 2 platform consists of three core models. The flagship Gemini Robotics 2 is a vision-language-action model that converts visual and language inputs into motor commands, enabling control of full humanoid robots from feet to fingertips. Gemini Robotics ER 2 serves as an embodied reasoning model that acts as the robot’s high-level planning brain, enabling multi-step task execution and human communication. Gemini Robotics On-Device 2 is optimized to run locally on robotic hardware without network connectivity.
According to the announcement, Gemini Robotics 2 enables humanoid robots like Apptronik’s Apollo 2 to execute complex whole-body tasks. For example, the model can understand instructions like “put the watering can into the green bin on the bottom shelf,” then walk to retrieve the object, navigate to the shelves, and place it precisely in the specified location. This represents the first time such models have controlled entire humanoid bodies rather than just upper-body movements.
The system also achieves advanced dexterity across different end effectors. It can control the 22 degree-of-freedom SharpaWave hand on the Apollo 2 robot to perform delicate actions such as tying knots or sealing ziplock bags, and can operate two-fingered parallel grippers on other platforms for complex packing tasks.
A notable feature is multi-robot collaboration, enabling different robot types to communicate and work together on tasks that a single robot could not complete alone. The embodied reasoning model can now execute longer task sequences lasting several minutes involving hundreds of decisions, with improved understanding of when tasks begin and end.
The On-Device model offers fast adaptation to new robotic bodies, requiring only a few hours of adaptation time with fewer than 200 examples, even for robots with drastically different shapes, sensors, and degrees of freedom.
Google emphasizes safety as foundational to the system. The company introduced ASIMOV-Agentic, a new benchmark for measuring agentic safety and uncertainty resolution. Gemini Robotics ER 2 includes enhanced safety features such as detecting nearby humans, triggering safety protocols, and bringing robots to safe stops when humans approach too closely.
Key facts
- Gemini Robotics 2 enables whole-body control of humanoid robots, including walking, crouching, and object manipulation
- The system can perform dexterous tasks like tying knots and sealing bags using multi-fingered hands or parallel grippers
- Multi-robot collaboration allows different robot types to communicate and complete complex workflows together
- Gemini Robotics On-Device 2 can adapt to new robot embodiments in just a few hours with minimal training data
- The embodied reasoning model can execute multi-step tasks lasting several minutes with improved task progress understanding
- Safety features include human detection, proximity monitoring, and proactive requests for human intervention when uncertain