Google DeepMind's Gemini Robotics 2 controls humanoid robots from feet to fingertips

Google DeepMind's Gemini Robotics 2 marks a leap in robot control, expanding from upper-body-only manipulation to full-body coordination. The model now drives humanoid robots from feet to fingertips, enabling actions like walking, crouching, stretching, and picking up objects. In demonstrations, Apptronik's Apollo 2 robot bends to grab a watering can and retrieves specific items from a shelf. The update also brings advanced dexterity, allowing control of complex five-fingered hands. That means robots can perform fine motor tasks such as sealing a Ziploc bag, tying a trash bag, or unscrewing a lightbulb—skills previously out of reach for most humanoids.

The companion model, Gemini Robotics ER (embodied reasoning) 2, improves multi-step task execution and now understands when tasks begin and end. It also enables heterogeneous robots to collaborate; one video shows Apollo 2 instructing Google's dual-arm robot to place tools in a bin during a garage cleanup. Google calls ER 2 its safest robotics model yet: it can detect nearby humans, trigger safety tool calls, and bring the robot to a safe stop if someone approaches too closely. Meanwhile, the on-device model, which runs locally without internet, adapts faster to new embodiments with drastically different shapes, sensors, and degrees of freedom. While movement speed still lags, this is a significant step toward real-world whole-body coordination.