Google DeepMind Unveils Gemini Robotics 2 for Whole-Body Robot Control
The suite moves beyond pre-programmed industrial actions by letting a single model control different robot shapes, from tabletop arms to full humanoids, while a separate model handles multi-step task planning.
Reporting from 1 source: GIGAZINE.
Google DeepMind announced Gemini Robotics 2 on July 30, 2026, a suite of Vision-Language-Action models that control robots from walking to fine two-handed tasks. The models handle posture, object manipulation, and multi-robot collaboration, adapting to different robot bodies. A higher-level model, Gemini Robotics ER 2, manages long task sequences and outperformed its predecessor in evaluations.
Google DeepMind's Gemini Robotics 2 is a Vision-Language-Action model that turns camera images and language instructions into robot actions. The same model can drive tabletop arms and full-body humanoids, combining arm movements with walking, crouching, and shifting the center of gravity.
Evaluations with the Apollo and Inspire hands showed success rates of 68.4 percent lifting objects from a table, 45.7 percent from the floor, and 76.3 percent from a shelf. Fine manipulation with two-finger grippers on the Franka Duo hit 74.2 percent for object transfer, 78.9 percent for packing tools, and 89.6 percent for part insertion. A five-fingered Sharpa hand with 22 degrees of freedom reached 92 percent removing a light bulb, though complex tasks like tying a garbage bag (44 percent) and closing a zippered bag (40 percent) remain difficult.
The companion Gemini Robotics ER 2 model plans long task sequences, making hundreds of decisions over several minutes. It outperformed the prior ER 1.6 across real-world VLA, simulated VLA, and human teleoperation control methods.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.