Google DeepMind has announced a significant update to its line of robotic AI solutions. At the center of attention is the Gemini Robotics 2 model, belonging to the VLA (Vision-Language-Action) class, which dramatically expands robot control capabilities, covering the entire body rather than just individual limbs. In parallel, an improved ER 2 system has been introduced, serving as a "high-level brain" for comprehensive task coordination.

From Manipulation to Full Control

The key difference between the new model and its predecessor is the transition from controlling the upper body to holistic control over the entire body. Gemini Robotics 2 now allows robots to perform complex dynamic actions: walking, squatting, bending, as well as manipulation in confined spaces. This is a critically important step for integrating robots into real, unstructured environments.

In demonstration tests, the Apptronik Apollo 2 humanoid, operating under the control of the new model, successfully handled tasks requiring fine motor skills: picking up a watering can, removing a baseball glove from a shelf, and locating target objects. Particularly impressive is the support for complex hands with 22 degrees of freedom, enabling delicate operations such as tying knots or unscrewing light bulbs.

ER 2: Orchestrating Multi-Robot Systems

The updated ER 2 model takes on the functions of a strategic planner. It is capable of decomposing multi-step tasks, monitoring their execution, and, when necessary, initiating a re-attempt. Moreover, ER 2 can distribute workloads among multiple robots simultaneously. For example, in one scenario, Apollo 2 coordinates the actions of a dual-arm robot to sort tools during cleanup.

Special attention deserves the local version, Gemini Robotics On-Device 2. This model operates fully autonomously, without cloud connections, and can be adapted to new types of robots in just a few hours of training based on fewer than 200 examples. This approach significantly lowers the barrier to entry for implementing AI in robotics startups.

Currently, Gemini Robotics 2 is available in private preview mode, while ER 2 has already been integrated into the Gemini API and Google AI Studio. On-Device 2 is undergoing testing with a limited circle of trusted partners.

My take: This is not just an incremental update, but a paradigm shift. The transition to "holistic intelligence" and multi-agent orchestration—these are precisely the directions that will determine the commercial viability of robotics in the coming years. The model's ability to adapt to new hardware platforms with a minimal number of examples is a key factor that could accelerate market consolidation and move humanoid robotics from the stage of pilot projects into the real sector of the economy.