.webp)
Google DeepMind has unveiled Gemini Robotics 2, a new system for humanoid and other robotic platforms.
The release centres on three models that can operate together or separately, covering action, reasoning and on-device adaptation for robotics tasks. Gemini Robotics ER 2 will be available on AI Studio, while the vision-language-action and on-device models are being offered to early-access partners.
Google DeepMind described Gemini Robotics 2 as a step towards more general-purpose physical artificial intelligence, focused on expanding the range of tasks robots can perform through whole-body movement and hand control. The system is designed for humanoids as well as other robot types, reflecting a wider industry push to build software that can transfer across different hardware.
Three models
At the core of the release is Gemini Robotics 2, a vision-language-action model that turns visual input and language prompts into physical actions. Alongside it is Gemini Robotics ER 2, an embodied reasoning model for task planning and human interaction, and Gemini Robotics On-Device 2, which runs locally on a robot and is intended to adapt more quickly to new embodiments.
This layered approach separates decision-making, action and local responsiveness within robotic systems. It also reflects a broader trend in robotics, as companies seek to combine centralised reasoning with hardware-specific control to make machines more flexible in real-world settings.
Body control
One of the main changes in the latest version is full-body control for humanoid robots. The model can control the entire machine from head to toe, allowing it to reason through movement as it walks, crouches and bends.
That matters because many robotics systems have historically focused on limited motion planning or narrow task execution, especially in controlled environments. Extending control across the full body is intended to improve balance, co-ordination and task completion in situations that require more natural movement.
Google DeepMind also highlighted more advanced dexterity in five-fingered robotic hands with 22 degrees of freedom. Examples included tying a trash bag and unscrewing a light bulb, tasks that rely on fine manipulation rather than simple gripping.
Improving dexterity remains a major challenge in robotics because hands must deal with changing object shapes, force levels and orientation. Progress in that area is often seen as central to moving robots beyond repetitive industrial work and into homes, warehouses and service roles, where tasks are less predictable.
Robot teamwork
Another part of the announcement is a system for multi-robot collaboration. The technology allows different types of robots, such as a humanoid and a bi-arm robot, to communicate and work together on tasks that would be harder for a single machine.
Co-ordinating more than one robot has become an increasingly important area of research as developers look to logistics, manufacturing and field operations. In those settings, dividing labour between different machines can improve efficiency, but only if software can manage communication and task allocation reliably.
Safety focus
Safety was another prominent theme in the release. Google DeepMind said it combines traditional physical safety measures with AI safety frameworks, and described Gemini Robotics ER 2 as its safest robotics model so far on the ASIMOV2 benchmark.
The company also introduced a new benchmark called ASIMOV-Agentic, which it said is designed for agentic physical AI and for navigating uncertainty. One example was testing whether a robot requests human intervention when it is uncertain.
That emphasis reflects growing scrutiny of AI systems that move and act in the physical world. While software errors in chatbots or search products can create misinformation or confusion, mistakes in robotics can lead to damaged goods, operational disruption or direct safety risks to nearby people.
The launch comes as major technology groups and specialist robotics companies race to build foundation models for machines that can interpret instructions, understand their surroundings and carry out tasks with less bespoke programming. Developers are increasingly betting that recent advances in large language models and multimodal AI can help robots generalise across more situations than earlier systems could.
For Google DeepMind, the announcement extends its push to apply Gemini models beyond text, images and code to embodied systems. Its focus on humanoids is notable because the category has become a focal point for investors and researchers seeking machines that can operate in spaces built for people without major infrastructure changes.
Availability remains selective for now, with the full stack not yet broadly open to all users. Gemini Robotics ER 2 will be accessible through AI Studio, while the action and on-device models are limited to early-access partners.
Google DeepMind said the release includes "3 highly capable models that work together or independently depending on the context".