Google DeepMind announced Gemini Robotics 2, an advancement in AI models designed to give robots intelligent control over their entire bodies, from feet to fingertips.
The system comprises three models: Gemini Robotics 2, a vision-language-action model that converts visual and language input into motor control for humanoid and bi-arm robots; Gemini Robotics ER 2, an embodied reasoning model that enables robots to plan multi-step tasks and communicate with humans; and Gemini Robotics On-Device 2, an efficient model optimized to run locally on robotic devices.
According to the announcement, Gemini Robotics 2 enables humanoid robots to perform whole-body tasks such as walking, crouching, and manipulating objects. The model can control fine dexterity tasks like tying knots or sealing ziplock bags using five-fingered hands, as well as complex manipulation with two-fingered grippers.
The embodied reasoning model allows robots to execute longer task sequences lasting several minutes and introduces multi-robot collaboration capabilities. Gemini Robotics On-Device 2 can adapt to new robot embodiments with just a few hours of adaptation time using fewer than 200 examples.
Google DeepMind stated that the models include safety enhancements, introducing ASIMOV-Agentic, a benchmark for agentic safety that measures the reasoning agent’s ability to refuse unsafe actions and request human intervention when uncertain. The embodied reasoning model is described as capable of detecting nearby humans and triggering safety measures.
Gemini Robotics ER 2 is available on Google AI Studio and in private preview on Gemini Enterprise Agent Platform, while the VLA and On-Device models are available to early-access partners.
Key facts
- Gemini Robotics 2 enables full-body control of humanoid robots including walking, crouching, and object manipulation
- The model can control five-fingered hands with 22 degrees of freedom to perform delicate tasks like tying knots
- Robots can adapt to new embodiments with few hours of adaptation time using fewer than 200 examples
- Multi-robot collaboration enables different robot types to work together on complex workflows
- Gemini Robotics On-Device 2 runs locally without requiring network connectivity
- The safety model ASIMOV-Agentic measures the agent’s ability to refuse unsafe actions and request human intervention