Google's new Gemini Robotics 2.0 gives robots better dexterity and real-time awareness

Google DeepMind has rolled out Gemini Robotics 2.0, an advancement that equips robots with the ability to tackle intricate maneuvers, assess their environment dynamically, and coordinate actions across multiple machines. A set of three new sub-models underpins the upgrade, with one now available for developer testing.

From party tricks to general-purpose machines#

Robot videos circulating online have long featured impressive feats—running, dancing, backflips—yet these typically represent pre-programmed routines for specific stunts. The vision behind Gemini robotics differs fundamentally: building a generalist robot capable of performing the broad range of tasks humans handle daily. Google DeepMind researchers sometimes refer to this aspiration as "physical AGI." The ideal is straightforward: instruct a robot on what to do, and it executes the command. The 2.0 release represents progress toward that goal, with enhanced control over whole-body movement and complex manipulation, including refined operation of intricate humanoid hands.

The embodied reasoning breakthrough#

The core innovation centers on Gemini Robotics ER 2, an updated "embodied reasoning" model that DeepMind positions as a substantial improvement over its 1.6 predecessor. This system is a vision language model (VLM) designed to parse both instructions and visual surroundings. The key enhancement: ER 2 processes live video streams from robot cameras in real time, enabling it to monitor task completion as the robot moves through its motions. Google reports that ER 2 achieves nearly 60 percent accuracy when assessing video frame completion—a meaningful leap beyond the 1.6 model and competing AI platforms' visual capabilities, though still short of flawless performance.

What's next#

The integration with the Gemini Live API opens the door for developers to explore this capability starting immediately.