Google DeepMind’s new AI model controls entire humanoid robots

Google DeepMind’s new Gemini Robotics 2 AI model can control a humanoid robot’s whole body, from its feet to its fingertips, the company announced on July 30, 2026.

The update builds on the earlier version, which handled only upper-body movements. Robots can now walk, crouch, stretch and grasp objects.

Videos released by the company show Apptronik’s Apollo 2 robot bending to pick up a watering can and taking items from a shelf.

The model also improves dexterity, with five-fingered hands capable of sealing a Ziploc bag, tying a trash bag or unscrewing a lightbulb.

Separately, Google DeepMind updated its Gemini Robotics ER 2 model, a vision-language system for embodied reasoning. It now better recognizes when tasks start and finish, enabling it to complete multi-step jobs over longer periods.

This lets different types of robots work together: one video features Apollo 2 instructing a dual-arm machine to put tools in a bin while cleaning a garage.

The company says ER 2 is its safest robotics model yet, with improved human detection that can trigger a safety stop if someone gets too close.

DeepMind also enhanced the on-device model that runs locally without an internet connection, allowing faster adaptation to robots with different shapes, sensors and degrees of freedom.