- Published on
Google Introduces Gemini Robotics 2 with ‘Whole Body Intelligence’
- Authors
- Name
- DeepHorus
Google DeepMind has unveiled Gemini Robotics 2, a family of AI models designed to provide robots with whole-body control, enhanced dexterity, multi-step reasoning, and collaborative capabilities. This release marks a significant step beyond narrow, preprogrammed robotic routines, enabling machines to interpret spoken or written instructions and translate them into coordinated movements across different robot bodies. The system aims to make robots more adaptable and intelligent in real-world environments.
The new model family comprises three distinct components. Gemini Robotics 2 is a vision-language-action model that converts visual and language input into motor commands. Gemini Robotics ER 2 serves as the reasoning layer, helping robots understand scenes, plan multi-step tasks, and interact with humans. Gemini Robotics On-Device 2 is designed to run locally on robot hardware, reducing dependence on internet connections and remote computing.
In tests, the system demonstrated impressive whole-body control using an Apptronik Apollo 2 humanoid robot. The robot walked, bent, and placed a watering can into a green bin on a lower shelf, showcasing integrated movement. However, performance varied: pick-up success rates were 68.4% from tables, 45.7% from floors, and 76.3% from shelves. Google DeepMind acknowledged that robot movement still needs to become faster and more reliable.
Gemini Robotics 2 also handles different hand types. On Apollo 2, it controlled a five-fingered SharpaWave hand with 22 degrees of freedom, achieving 92% success unscrewing a light bulb but only 36% screwing it in. Other tasks like tying a trash bag (44%) and sealing a zip-close bag (40%) had lower success. In contrast, two-finger grippers on a Franka Duo robot performed better, with 74.2% on general pick-and-place, 78.9% on tool kitting, and 89.6% on precise insertion.
The reasoning layer, Gemini Robotics ER 2, enables longer task sequences lasting several minutes and involving hundreds of decisions. It can adjust when an action fails and track progress by understanding when tasks begin and end. The update also introduces multi-robot collaboration, allowing different machines to divide workflows based on their capabilities.
Local processing through Gemini Robotics On-Device 2 is crucial for environments where internet service is unavailable or unreliable. The model can adapt to new two-arm robot designs with fewer than 200 examples, making deployment more flexible. Additionally, the ASIMOV-Agentic benchmark tests whether robots can reject unsafe actions and request human help, addressing safety concerns.
Industry observers view this development as a practical step toward more versatile robots for manufacturing, logistics, and home assistance. By combining whole-body control with strong reasoning and on-device processing, Google DeepMind is addressing key limitations in current robotics. The focus on collaboration and safety benchmarks also signals a maturation of the field.
Looking ahead, further improvements in speed and reliability are expected, particularly for complex manipulations. As the models continue to learn from real-world data, robots may soon handle a wider range of tasks with greater autonomy. The introduction of Gemini Robotics 2 could accelerate the adoption of intelligent robots across industries, reshaping how machines interact with their environments.
