Google DeepMind Unveils Gemini Robotics 2 AI for Humanoids
Google DeepMind is teaching steel skeletons how to walk, think, and handle chores, though domestic independence is still noticeably far from reality.
The new release introduces three specialized models under the Google DeepMind banner, designed to coordinate complex physical actions. The primary Gemini Robotics 2 vision-language-action model now controls full-body motions simultaneously, managing walking, bending, and balance alongside arm movements.
During demonstrations on the Apollo 2 robot created by Apptronik, the system received instructions to stash a ruler on a low shelf, successfully navigating across the room and squatting down to complete the task at a painfully leisurely pace. High-precision dexterous hands with 22 degrees of freedom can tie plastic bags or insert small components, yet screwing in a lightbulb succeeds only 36% of the time despite a 92% success rate at unscrewing it.
High-level task planning is driven by Gemini Robotics ER 2, built on top of Gemini 3.5 Flash to process continuous real-time video, spot physical errors, and coordinate multi-robot operations. While this brain model is accessible via public API, the low-level movement controls remain strictly locked to select partners like Boston Dynamics and Agile Robots.
Humanity stands at a fascinating crossroads where artificial intelligence can reason across 128,000 tokens of continuous video data, yet leaves homes pitch-black nearly two-thirds of the time when attempting standard maintenance.
Source: Google DeepMind
Comments
This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.