Google DeepMind Unveils Gemini Robotics 2: Full-Body Control and Dexterous Hands

Google DeepMind has released Gemini Robotics 2, the latest version of its vision-language-action model for robots, on August 2. The new model upgrades from upper-body tabletop tasks to intelligent whole-body control, advanced dexterity, and multi-robot collaboration.

In a demonstration with Apptronik Apollo 2, the robot followed a complex instruction to fetch a watering can and place it into a green bucket on a lower shelf. It autonomously walked to the table, picked up the can, navigated to the shelf, and placed it correctly, requiring coordinated decision-making and balance throughout.

The model now controls a five-fingered, 22-degree-of-freedom dexterous hand on Apollo 2, enabling fine actions like tying knots or sealing bags. It also works with standard two-finger parallel grippers on Franka Duo for tasks such as compact packing.

Gemini Robotics 2 introduces multi-robot collaboration, allowing different robot types to communicate and work together on complex workflows. The ER 2 model acts as a high-level brain, identifying task start and end points, detecting key events, and self-correcting if a step fails.

DeepMind also launched the ASIMOV-Agentic benchmark to measure a robot's ability to reject unsafe tool calls, predict task feasibility, and request human intervention when uncertain. The ER 2 model performed best in safety constraint and human proximity tests, detecting nearby people and triggering safe stops.

For factory and warehouse environments requiring no network latency, DeepMind introduced Gemini Robotics On-Device 2, which runs locally. It can adapt to new dual-arm robot morphologies in under 200 data samples and a few hours, even with different shapes, sensors, and degrees of freedom.

From upper-body control to whole-body coordination, and from following commands to deciding when to stop, Gemini Robotics 2 brings robots closer to everyday life. The ability to tie knots and seal bags with a dexterous hand hints at tasks like tying shoelaces, pushing robots further into real-world applications.