Google DeepMind has introduced Gemini Robotics 2, an advanced AI system designed to give robots intelligent whole-body control and reasoning capabilities. According to the announcement, the system represents a major advance toward robots that can adapt to unpredictable environments and complete complex real-world tasks.
Most existing robots rely on pre-programming or teleoperation for narrow, repetitive tasks and struggle to transfer learned skills between different robot bodies. Gemini Robotics 2 addresses this by enabling robots to reason through movements and adapt across different physical forms.
The system comprises three models. The primary Gemini Robotics 2 is a vision-language-action (VLA) model that converts visual and language inputs into motor control, enabling full humanoid control from feet to fingertips, including dexterous manipulation. Gemini Robotics ER 2 serves as an embodied reasoning model, allowing robots to understand their physical environment, plan multi-step tasks lasting several minutes, and communicate with humans. Gemini Robotics On-Device 2 is optimized for running locally on robots without network connectivity.
According to the source, the system demonstrates whole-body coordination capabilities. In one example, when instructed to “put the watering can into the green bin in the bottom shelf,” Apptronik’s Apollo 2 humanoid robot can walk to a table, pick up the object, navigate to shelves, and place it in the correct location. The system also enables advanced dexterity, controlling a 22 degree-of-freedom hand to perform delicate actions like tying knots or sealing ziplock bags, as well as operating two-fingered grippers for complex packing tasks.
A new feature is multi-robot collaboration, enabling different robot types to communicate and work together on complex workflows that single robots cannot accomplish alone. The reasoning model can now execute task sequences lasting several minutes involving hundreds of decisions, with improved ability to understand when tasks begin and end.
For on-device operation, Gemini Robotics On-Device 2 can adapt to new robot designs in just a few hours of adaptation time, typically requiring fewer than 200 examples. This capability works across robots with different shapes, sensors, and degrees of freedom.
Google has also introduced ASIMOV-Agentic, a new safety benchmark measuring how well the system refuses unsafe actions and requests human intervention when uncertain. According to the announcement, Gemini Robotics ER 2 represents their safest robotics model to date in safety constraint following and can detect nearby humans to trigger safe stops.
Key facts
- Gemini Robotics 2 enables full-body control of humanoid robots, expanding beyond previous upper-body-only capabilities
- The system can execute complex multi-step tasks lasting several minutes with hundreds of decisions
- Multiple robots can now collaborate together to solve workflows a single robot cannot complete alone
- On-device models can adapt to entirely new robot embodiments in just a few hours with fewer than 200 examples
- Gemini Robotics ER 2 can detect human proximity and trigger safety stops when humans approach too closely