Google DeepMind launches two AI models, Gemini 2.0-based Robotics and Robotics-ER, to help robots “perform a wider range of real-world tasks than ever before”
Gemini Robotics also makes robots more dexterous, allowing them to perform more precise tasks, like folding a piece of paper.
Context & Ripple Effects
This launch moves Gemini from general-purpose multimodal reasoning toward robot control. It builds on DeepMind’s earlier work using Gemini 1.5 Pro to guide robots from simple instructions, but adds a dedicated robotics model and a companion model for embodied reasoning.
The product line subsequently expanded to 1.5 models aimed at multi-step robot work and an on-device version with an SDK, making this initial release the starting point for a broader effort to turn Gemini capabilities into deployable robot software.
First-order effects
- Google DeepMind gains two Gemini 2.0-based models tailored to robot use: Gemini Robotics for action and dexterity, and Robotics-ER for reasoning about the physical world.
- Robot developers working with DeepMind’s stack can target finer manipulation tasks, such as paper folding, rather than relying solely on high-level language understanding.
Second-order effects
- The split between action and embodied-reasoning models raises the bar for robotics AI rivals: competitive offerings must improve both physical-task reliability and the reasoning that precedes an action.
- The later move to an on-device robotics model and developer SDK suggests the initial models also create a path for developers to move from cloud-centered demonstrations toward product integration.
Third-order effects
- If model families continue to advance from reasoning to action to local deployment, robotics competition will increasingly center on reusable foundation-model stacks rather than one-off task programming.
- That shift could favor AI providers able to pair general models with developer tooling and robot-specific adaptation, though real-world safety and reliability remain the limiting test.
The trend: This is an early point in the rise of embodied AI, in which frontier-model providers package perception, reasoning, and action capabilities for robots.