Google DeepMind’s Gemini Robotics 2 Gives Robots Whole-Body Intelligence and Dexterity

Google DeepMind has taken a significant leap in robotics AI with the release of Gemini Robotics 2 on July 30, 2026. This next-generation model suite extends control from mere upper-body manipulation to full, intelligent whole-body coordination, enabling robots to walk, crouch, stretch, and perform complex tasks with newfound dexterity. This marks a pivotal step toward what DeepMind calls "physical AGI."

What is Gemini Robotics 2?

Unlike earlier systems that focused on pre-programmed or teleoperated sequences, Gemini Robotics 2 is a suite of three distinct AI models designed to give robots the ability to think, act, and adapt. The core innovation is the integration of vision, language, and motor control, allowing a single AI to manage everything from a humanoid's gait to the fine manipulation of a five-fingered hand.

The system comprises a Vision-Language-Action (VLA) model for direct motor control, an Embodied Reasoning (ER) model for high-level task planning and human interaction, and an On-Device VLA optimized for local, low-latency operation. This trio allows robots to understand complex instructions, reason about their environment, and execute tasks in real-time.

Whole-Body Control and Advanced Dexterity

The most striking advancement is the ability to control entire humanoid robots. In a demonstration, DeepMind showed the Apptronik Apollo 2 humanoid processing the instruction to "put the watering can into the green bin on the bottom shelf." The robot then autonomously walked to a table, picked up the can, navigated to a shelf, and precisely placed it by bending its knees and waist.

This whole-body coordination extends to advanced dexterity. The model can control the 22-degree-of-freedom SharpaWave hand to perform delicate actions like tying knots or sealing a ziplock bag. It also handles standard two-fingered grippers on platforms like the Franka Duo for complex tasks such as tight packing. While precision and speed are still being refined, the results are impressive: 76.3% success rate in shelf retrieval, 68.4% in tabletop pick-up, and 45.7% in floor lifting.

Multi-Robot Collaboration and Fast Adaptation

Gemini Robotics 2 introduces multi-robot collaboration, allowing different types of robots to communicate and work together on complex workflows that a single robot could not handle alone. This is powered by the Embodied Reasoning model, which serves as a high-level brain, coordinating tasks and tracking progress over several minutes.

Perhaps most importantly for real-world deployment, the On-Device model can adapt to entirely new robot embodiments with fewer than 200 examples and just a few hours of additional training. This is a radical departure from the traditional approach of training separate AI for each robot, paving the way for a single, versatile intelligence layer across diverse hardware.

Safety and the Path to Physical AGI

Safety is a core focus of the release. DeepMind introduced the ASIMOV-Agentic benchmark to measure a robot's ability to refuse unsafe actions and request human intervention when uncertain. The ER 2 model also features enhanced human proximity detection, enabling robots to safely stop when a person approaches too closely.

"It's another milestone in our path towards really getting towards what we call physical AGI, which means we get a robot to do anything that a human can," said Carolina Parada, head of robotics at Google DeepMind, in an interview with Wired. The release solidifies Google DeepMind's leadership in embodied AI, moving beyond the digital realm to tackle the hardest problems in the physical world.