Google DeepMind has released Gemini Robotics 2, the latest version of its vision-language-action model embodied in a humanoid robot.

In a blog post, the company said the launch is a significant step toward “physical AGI” — artificial general intelligence that can interact with and operate in the real world through physical systems such as robots, allowing robots to perform physical tasks that humans can do.

By combining multiple AI models, the system enables what DeepMind calls “intelligent whole-body control.” The company said this gives robots the ability to perceive their surroundings, reason through multi-step tasks, and coordinate movement across their entire bodies. Robots integrated with the system can also collaborate with other robots, underscoring how robotics systems are increasingly being built around coordination rather than single-task automation.

DeepMind’s robotics lead, Carolina Parada, has described the long-term ambition as enabling “a robot to perform any task a human can.”

Video demonstrations released last week showed Apptronik’s Apollo 2 robot and robotic hands from Sharpa autonomously completing household and industrial tasks such as tying trash bags and screwing in lightbulbs. The systems were trained using a combination of teleoperation, video demonstrations and simulation.

Paul Miller, an analyst at Forrester, said the launch builds on Google’s earlier robotics work by improving collaboration between robots, helping machines recover from failed actions and giving the model a better understanding of an entire robot rather than focusing only on individual components such as an arm or gripper.

Still, Miller said the leap to bridge the gap between the robotic brain and body carries significant risks.

“Robots are physical machines,” he said. “They may be strong, and they may be heavy. People working with these machines, and people encountering them out in the world, need to be confident that these machines are safe.”

He pointed to challenges including ensuring robots fail safely if sensors or power systems malfunction, arguing that widespread deployment alongside humans will depend on robust safety safeguards.

“Before robots can really be deployed alongside people, the safety case must be convincingly and believably proven,” he said.

On Google’s goal of reaching physical AGI, Miller said the promise remains “aspirational” for now.

“A robot that can eventually do anything a human can, in any environment, may be a reasonable research goal, but it’s one we’re nowhere near,” he said.

For now, he expects physical AI to deliver value in controlled environments such as factories and warehouses, where tighter layouts and more predictable tasks make it easier to limit risk and uncertainty.

“Physical AI may be adaptable and more flexible than previous forms of automation, but it’s still easier and cheaper to deploy automation in environments where you can minimize risk, minimize uncertainty, and maximize utilization,” he said. “A factory or warehouse may be noisy, dirty, or chaotic, but it’s an awful lot simpler for a robot than a domestic environment.”