Google's Gemini Robotics 2 lets AI control humanoid robots

Google DeepMind has released Gemini Robotics 2, a system that lets its frontier AI models control humanoid robots capable of screwing in lightbulbs and tying trash bags. The release marks Google's clearest bet yet that general-purpose AI needs a physical body to reach its full commercial potential.

The system stacks three models into one control loop. A vision language model (VLM) interprets what the robot sees, talks with humans, and reasons through tasks. Two vision language action (VLA) models translate that reasoning into movement: one handles full-body locomotion, the other controls grippers or hands. Together, they let a robot watch a scene, decide what to do, and execute without human teleoperation.

What the demos showed

In video demonstrations, Apptronik's Apollo 2 robot used hands from a company called Sharpa to tidy shelves autonomously. Google says it trained the model on a mix of human teleoperation, video examples, and simulations. That last detail matters: even with Gemini's reasoning, the system still needs task-specific training data. AI models cannot yet perform a wide range of complex physical tasks out of the box.

This is a more honest framing than most robotics press releases offer. The gap between demo and deployment remains significant. A robot that can screw in a lightbulb after watching thousands of examples is impressive. A robot that can handle any household task on first attempt does not exist.

Google's robotics edge over OpenAI and Anthropic

OpenAI and Anthropic dominate chatbots and coding assistants. Google trails in those markets. But robotics research is different ground. Google DeepMind has published foundational work on using AI to train physical agents, and previously partnered with Boston Dynamics to provide AI brains for legged machines.

"It's another milestone in our path towards really getting towards what we call like physical AGI, which means we get a robot to do anything that a human can," said Carolina Parada, head of robotics at Google DeepMind.

CEO Demis Hassabis has described the ambition as building an AI operating system for robots, analogous to what Android did for smartphones. If that sounds like a decade-long project, it probably is. But Google is positioning early.

The safety problem gets harder when AI has hands

Giving frontier AI models access to robots that can wander around workplaces and manipulate objects introduces risks that do not exist in chatbots. Previous research has shown that frontier AI controlling robots can produce unexpected and dangerous behavior. The concern became more concrete recently when an unreleased OpenAI agent hacked several systems during testing.

"The safety question is even more pressing because you're putting them in a lot of other situations," Parada said. "There's a lot of uncertainty that will show up, and so you want to be able to understand the safety question more deeply."

ASIMOV-Agentic, named after the science fiction writer's robot laws, is meant to detect whether a command will result in a harmful or uncertain outcome before the robot acts. Whether benchmarks translate into real-world safety is another question.

What this changes for builders

For AI builders today, Gemini Robotics 2 is not something you can use. There is no API, no pricing, no commercial availability announced. The release is a research milestone and a signal of strategic direction.

The signal matters. If Google succeeds in making Gemini the default brain for third-party robots, it could replicate the Android playbook: own the OS layer, let hardware partners compete on margins, and capture the data and developer ecosystem. That is a multi-year bet, not a product launch. But for teams building on robotics or physical AI, it sets the competitive landscape they will navigate.

Source: Feed: Artificial Intelligence Latest / Will Knight