Google DeepMind continues to set the pace in robotics, announcing Gemini Robotics 2 — an innovative VLA model (Vision-Language-Action) that translates text and visual data directly into motor actions. This is not just an update, but a qualitative leap: the system now covers control of the robot's entire body, not just its upper part, as in the previous version.

From manipulators to humanoids: versatility without compromise

The model is designed for robots of various form factors — from compact desktop manipulators to full-sized humanoids. The key difference is the ability to perform complex movements: walking, squatting, bending, and picking up objects in confined spaces. In demonstration tests, the Apptronik Apollo 2 humanoid, powered by Gemini Robotics 2, confidently handled tasks like picking up a watering can or removing a baseball glove from a shelf, showcasing impressive coordination.

Improved motor skills deserve special attention. The model supports complex hands with 22 degrees of freedom, opening up possibilities for fine operations: tying knots, closing bags, or unscrewing light bulbs. This brings robots closer to human dexterity, which is critical for real-world industrial and household scenarios.

ER 2: a "high-level brain" for coordination

In parallel, DeepMind introduced the updated ER 2 model, which serves as a strategic planner. It breaks down multi-step tasks into stages, tracks their execution, and distributes the workload among multiple robots. In one of the showcased scenarios, Apollo 2 coordinated the actions of a Google dual-arm robot that was placing tools into a container during a garage cleanup. This is a clear example of seamless collaboration between the "brain" and "body" in multi-robotic systems.

For those who value autonomy, the company offered a local version, Gemini Robotics On-Device 2. It operates without an internet connection and adapts to new robots in fewer than 200 examples, learning in just a few hours. This makes the technology accessible for specialized industrial applications where data security and response speed are of paramount importance.

Currently, Gemini Robotics 2 is available in private preview, ER 2 via the Gemini API and Google AI Studio, and On-Device 2 is being tested by a limited circle of trusted partners. Clearly, DeepMind is purposefully moving toward creating universal systems capable of integrating into the real economy.

My analysis: Developments like these signal a transition from experimental robotics to practical solutions. However, the key challenge will remain not so much technical power as the economic viability of deploying such systems in mass production. In the next 2-3 years, we will likely see the first commercial pilots where these technologies will be stress-tested in real-world conditions.