Google DeepMind’s Gemini Robotics 2 Brings AI Closer to Humanoid Robots

Google DeepMind has unveiled Gemini Robotics 2, a new artificial intelligence system that can control multiple kinds of robots, including humanoid machines performing delicate real-world tasks such as screwing in lightbulbs and tying trash bags. The launch signals Google’s most ambitious push yet to move Gemini beyond screens and into physical environments, where the stakes for safety, reliability and usefulness are much higher.

The release matters because it moves one of the world’s most capable AI efforts closer to what Google executives describe as a long-term goal: software that can help robots operate in homes, warehouses and workplaces with far less task-specific engineering. It also intensifies the race among major AI companies to prove that their models can do more than converse, code or search; they must also sense, plan and act in the physical world.

DeepMind’s latest robotics system blends several models into a single stack. In practice, that means the robot can interpret visual input, understand instructions in natural language and translate those decisions into movement across a body, arms, hands and grippers. The company says the approach is designed to help robots reason about their surroundings, interact with objects and carry out more complex jobs autonomously.

What Google DeepMind announced

Google DeepMind says Gemini Robotics 2 combines a vision-language model with two vision-language-action models so a robot can both understand what it sees and decide how to move. The system is meant to unify perception, reasoning and control rather than treat them as separate robotics problems.

The first component, a vision-language model, processes images and video, interprets instructions and helps a robot reason about what to do next. The two action models then handle physical motion: one for broader body movement and another for hands or grippers. Together, the models give robots a more integrated decision-making pipeline than earlier systems that often relied on narrower task-specific routines.

Google shared demonstrations showing different robots performing a variety of chores on their own. In one example, Apptronik’s Apollo 2 humanoid worked with hands made by Sharpa to organize shelves. In other examples, the company showed humanoid and non-humanoid platforms completing tasks that require coordination, object handling and spatial awareness.

How Gemini Robotics 2 works

Gemini Robotics 2 works by linking perception and action through a layered model architecture. The AI system first interprets the scene, then decides what action should happen, and finally converts that decision into movement commands for the robot’s body parts.

That approach is important because robotics has long been constrained by a gap between AI that understands language and AI that can physically act. A chatbot can explain how to tie a knot; a robot must also know where the rope is, how tightly to grip it, and how to compensate when the object slips or shifts.

DeepMind says the system was trained using a mixture of human teleoperation, video examples and simulation. That training recipe remains necessary because frontier models still cannot reliably generalize to a wide range of physical tasks without examples of those tasks or similar ones.

• Teleoperation gives the model examples from human-guided robot actions.
• Video training helps it learn patterns from visual demonstrations.
• Simulation provides a safer, cheaper environment for practice and iteration.

In other words, the model is not a magic general-purpose robot brain. It is an advanced system that still depends on structured preparation, carefully defined tasks and substantial training data before it can perform useful work.

Why this release matters for the AI race

This release matters because Google is signaling that the future of AI is not limited to software products like chatbots, coding assistants or search tools. The company is increasingly positioning Gemini as the foundation for systems that can cross from digital output into physical action.

That strategy sets Google apart in emphasis, even if it is not alone in the market. Anthropic and OpenAI have drawn much of the recent attention for chatbots and AI coding tools, but Google still has one of the deepest histories in robotics research. DeepMind has published important work on robot learning, and Google previously worked with Boston Dynamics on bringing AI capabilities to legged robots.

The larger bet is that a model family like Gemini may eventually function as a common intelligence layer across many robot designs, the way Android operates across a broad range of smartphones. Google CEO Demis Hassabis has described that ambition in those terms: a platform for multiple robot types rather than a single machine.

Carolina Parada, who leads robotics at Google DeepMind, described the launch as a step toward what the team calls “physical AGI,” referring to a system that could eventually enable a robot to do the kinds of tasks a human can.

That phrase is still aspirational, but it captures the scale of Google’s ambition. The company is not merely optimizing a warehouse robot or a home assistant. It is trying to build a general robotics foundation model that could adapt across settings and embodiments.

What kinds of robots can Gemini control?

Gemini Robotics 2 is designed to work across a range of robot types rather than only one form factor. That includes humanoids, which are especially useful for tasks built around human spaces and human tools, as well as other robotic systems with different body structures and end-effectors.

Humanoid platforms have attracted intense interest because they can, in theory, use the same rooms, shelves, handles and tools already built for people. A robot that can stand upright, reach with arms and manipulate objects with dexterity has a broader immediate use case than many specialized industrial machines.

At the same time, humanoids are technically difficult. Coordinating balance, perception, grasping and manipulation in a single model is much harder than solving only one of those problems. That is why Google’s multi-model architecture is notable: it reflects an attempt to make robots more capable without asking one model to handle every part of the problem in isolation.

Examples shown in the demos

Google’s preview videos highlighted several practical, if still controlled, tasks. The company showed robots navigating the kind of environments where real work happens: shelves, objects, containers and household items that require careful handling rather than repetitive industrial motion.

Among the tasks demonstrated were:

• tidying shelves
• screwing in lightbulbs
• tying trash bags
• handling objects with grippers and hands

These examples are modest compared with the dream of fully autonomous household robots, but they matter because they show the direction of travel. The industry is moving from brittle single-purpose systems toward AI agents that can perceive and act in a more flexible way.

How was Gemini Robotics 2 trained?

Gemini Robotics 2 was trained using human teleoperation, video examples and simulation, according to Google DeepMind. The company has not suggested that the model can simply infer every new physical task on its own without examples.

That distinction is critical. In software, large language models can often generalize from vast amounts of text. In robotics, the physical world introduces friction, uncertainty and non-repeatable conditions. Objects can be heavier than expected, surfaces can be slippery, and a slight misalignment can cause failure.

Training with human demonstrations gives the model concrete examples of successful action. Simulation expands the number of scenarios the system can practice in without risking hardware damage. Video examples add another layer of observational learning that helps the AI map visual patterns to likely actions.

Even so, the model still appears to require task-specific guidance. That is one reason robotics has remained so challenging: intelligence in the abstract is not the same as dependable movement in the real world.

Why safety is now central to robot AI

Safety is becoming the defining issue as frontier AI models gain access to robots that can move through homes, factories and other environments. A chatbot may generate a wrong answer, but a robot driven by a flawed model can create physical harm or property damage.

That concern is not theoretical. Research has already shown that controlling robots with advanced AI can produce unpredictable behavior, sometimes in ways that become dangerous. The more open-ended the setting, the more difficult it becomes to anticipate every possible failure mode.

Google’s concern is heightened by what has happened in the digital world, where AI agents have at times made unintended decisions with real consequences. OpenAI recently disclosed that an unreleased agent had been able to break into multiple systems, underscoring how quickly autonomous AI can behave in ways developers did not expect.

Parada said the safety challenge becomes more urgent when AI is placed into real-world settings, because robots encounter more uncertainty and more situations the system has not previously seen.

For Google, the answer is not to slow down robot development but to build multiple layers of protection into the stack. The company says guardrails are applied at each model layer so that errors or unsafe recommendations can be caught before becoming physical action.

What is ASIMOV-Agentic?

ASIMOV-Agentic is Google DeepMind’s new benchmark for testing whether multiple AI systems working together to control a robot can make unsafe or uncertain decisions. The benchmark is designed to detect when a command might lead to harmful outcomes before the robot acts.

Benchmarking has become a crucial part of the AI industry because companies need objective ways to compare not just performance but reliability. In robotics, that need is even greater, because a model that scores well in a controlled test can still fail when it encounters clutter, ambiguity or unexpected human behavior.

Google’s move to introduce a dedicated safety benchmark suggests the company recognizes that robotics will not scale unless safety evaluation becomes as important as raw capability.

How does Google’s strategy compare with OpenAI and Anthropic?

Google’s strategy differs from its better-known AI rivals by emphasizing embodiment. OpenAI and Anthropic have captured much of the public attention with chatbots and coding assistants, but Google is using Gemini to push into systems that can physically manipulate the world.

That does not mean the company is ignoring conversational AI. Rather, it is trying to extend the same underlying intelligence stack into new product categories where the payoff could be larger and the competition thinner. If successful, robot intelligence could become a platform business as important as search or mobile software.

Google also enters the field with an advantage in research depth. Its work in robotics predates the current generative AI boom, and DeepMind has long explored how machine learning can improve manipulation, planning and control. The question is whether that research leadership can translate into practical systems robust enough for commercial deployment.

The answer will depend on several hard problems:

1. Whether robots can generalize beyond the tasks used in training.
2. Whether safety systems can prevent damaging or dangerous actions.
3. Whether the hardware is affordable and reliable enough for real use.
4. Whether enterprises and consumers trust AI-driven robots in shared spaces.

What comes next for Google’s robotics push?

What comes next is likely to be more testing, more benchmarks and a gradual move from demo videos to more constrained real-world deployments. Google has not suggested that Gemini Robotics 2 is a finished product ready to roam homes freely. Instead, it appears to be a platform release intended to push the state of the art and establish the company as a leader in robot foundation models.

The near-term challenge is to prove that the system can work repeatedly, safely and across different robot forms. The longer-term challenge is to determine whether one AI foundation can really scale across factories, warehouses and domestic settings without heavy customization.

If Google succeeds, Gemini could become a kind of operating system for robots: a shared intelligence layer that different hardware makers can adopt and adapt. If it falls short, the release will still mark an important moment in the industry’s move toward embodied AI, where language, vision and movement are increasingly fused into one system.

For now, the announcement confirms that the battle over AI’s future is no longer only about who builds the most persuasive chatbot or the fastest coding assistant. It is also about who can give machines the ability to understand the physical world, act on it safely and deliver useful labor at scale.

Timeline: Google’s robot AI push

The company’s latest announcement fits into a broader sequence of efforts to connect AI research with robotics hardware.

That progression shows a company trying to turn research credibility into a practical robotics platform. The outcome could shape not only how robots are built, but how quickly AI leaves the screen and enters the real world.