Google DeepMind launched Gemini Robotics 2.0 this month, a robot-control system designed to provide whole-body control for humanoid robots [1].
This development represents a shift toward more fluid and safe interaction between humanoid machines and their environments. By integrating an AI "brain" capable of managing complex physical movements, the system aims to make robots more viable for practical, real-world tasks that require high precision.
The new framework focuses on improving dexterity and safety. The company said Gemini Robotics 2.0 enables robots to perform complex tasks with greater ease while adhering to a new safety benchmark [1, 2]. This benchmark is intended to support multi-robot coordination, ensuring that multiple units can operate in the same space without compromising safety [1].
DeepMind developed the system to bridge the gap between high-level AI reasoning and the physical execution of movements. The whole-body control allows the AI to coordinate various joints and limbs simultaneously, a necessary step for humanoid robots to maintain balance while manipulating objects [3, 5].
While the announcement introduces the broader capabilities of the system, the rollout is gradual. Gemini Robotics 2 includes three distinct models, but only one is publicly available at this time [1]. The remaining models are expected to be released as the technology matures and the safety benchmarks are further validated.
Market analysts and prediction platforms are monitoring the trajectory of Google's AI development. One report said a prediction that Google will have the best AI model by December 2026, with an 11% probability [6].
“Gemini Robotics 2.0 enables full-body control for humanoid robots.”
The transition from limb-specific control to whole-body coordination is a critical hurdle in robotics. By implementing a standardized safety benchmark for multi-robot coordination, Google is attempting to move humanoid robots out of isolated lab settings and into collaborative environments. The limited public release of only one of the three models suggests that while the core architecture is ready, the most advanced or specialized versions of the AI require more rigorous safety testing before widespread deployment.
