Google Deepmind released Gemini Robotics 2, a vision-language-action model designed to control robots across different form factors, from compact tabletop arms to full-body humanoids. The model represents the company's most advanced robotics AI system to date.

Gemini Robotics 2 processes visual input and natural language instructions to generate motor commands for robot bodies. This unified approach allows a single model to work across diverse hardware without requiring task-specific retraining. The system handles both fine-grained manipulation tasks and whole-body coordination challenges.

The release also includes Gemini Robotics ER 2, which adds a higher-level reasoning layer on top of the base model. This reasoning component breaks down complex robotics tasks into executable steps, improving performance on multi-step objectives and decision-making scenarios.

The appeal of a generalist robotics model lies in deployment efficiency. Rather than training separate systems for tabletop manipulation, mobile bases, or humanoid platforms, operators can rely on one foundation model adapted through prompting or lightweight fine-tuning. This reduces development cycles and potentially cuts costs for robotics companies integrating the technology.

Google Deepmind's approach aligns with industry trends toward scaling language models beyond text and images into embodied AI. Companies including Boston Dynamics, Figure, and Tesla are pursuing similar paths. The technical challenge remains acute: robots operate in physical space with real consequences for errors, making robustness and reliability non-negotiable requirements.

Gemini Robotics 2's release suggests Google views robotics as a core frontier for AI scaling. The company has invested heavily in robotics research through Deepmind acquisitions and internal teams. Whether the model achieves widespread adoption depends on its real-world performance, integration complexity, and competition from alternative approaches like behavior cloning or reinforcement learning systems tailored for specific robot morph