Lead: Google DeepMind is driving a paradigm shift in embodied AI. The new Gemini Robotics 2.0 model lineup teaches machines not just to walk, but to understand video feeds, coordinate actions, and work in teams—transforming humanoids from programmable mannequins into autonomous agents.
Gemini Robotics ER 2: From Isolated Tasks to Orchestration
Google DeepMind has announced Gemini Robotics ER 2—a model that provides robots with advanced video understanding, tool orchestration, and multi-robot cooperation. Previously, robots solved narrow tasks in strictly defined conditions. Now, the focus has shifted to reasoning and interacting with the real world via visual context.
In my view, this is a pivotal step toward creating true physical agents. The model's ability to act as a link between multiple robots means we are approaching swarm behavior in industry and logistics. Instead of a single humanoid overloaded with computations, we get a coordinated ensemble where each agent understands its role in the overall process.
Gemini Robotics 2.0: "Embodied Intelligence" and Dexterity
The second version of the platform, Gemini Robotics 2, introduces the concept of whole body intelligence. This means the AI manages the robot's entire body as a single system rather than controlling individual joints, radically improving dexterity and manipulation safety.
The platform includes three models, but only one is currently available publicly. This suggests that the commercialization of complex motor-cognitive systems is proceeding cautiously. Safety in the physical world is a critical factor. An LLM error in a chat costs a user some frustration; an error by an 80-kilogram robot in a factory can lead to injuries and destruction. Google clearly recognizes this difference.
LLM Vulnerabilities: A Brake for Autonomous Agents
While robotics is leaping forward, the foundation it is built upon is showing structural problems. Researchers at the International Conference on Machine Learning (ICML) presented work proving that LLMs cannot be made completely secure from attacks due to a fundamental flaw in their architecture.
For embodied AI, this is a critical warning. If a language model serves as the "brain" for a physical robot, its vulnerability to prompt injections or hacking becomes a direct threat to physical safety. In my opinion, until the security of base models is solved at the architectural level, deploying autonomous robots into unstructured environments will remain risky. Robotics developers will have to build isolated safety layers rather than relying on the "intelligence" of the LLM itself.
Summary
Robotics is shifting from an obsession with mechanics to a focus on cognitive coordination. In the coming months, we should expect the integration of Gemini Robotics 2.0 into pilot industrial projects, where the ability of multiple robots to cooperate will yield real economic effects. However, the speed of adoption will be limited not by engineering capabilities, but by the security of the underlying AI models, whose vulnerability to attacks remains an unsolved systemic problem.