Google DeepMind's Gemini Robotics 2 Extends Control From the Waist Down: Walking, Crouching and Manipulating in One Policy
DeepMind announced Gemini Robotics 2 on 2026-07-30 in three variants — the VLA, an ER 2 embodied-reasoning model for multi-step planning, and an On-Device 2 build. The step past 1.5 is whole-body locomotion control (previously upper-body only), mapping vision and language to motor commands 'from feet to fingertips' so a humanoid can walk, crouch, reach and manipulate as one action. Reported ranges: 45.7–76.3% on general whole-body manipulation across shelf/table/floor, 74.2–89.6% on gripper pick-place and insertion, and a wide 32–92% on multi-finger tasks; platforms include Apptronik Apollo 2, Franka Duo, Dexmate, SO101 and Trossen, with ER 2 in private preview on AI Studio and the VLA/on-device models gated behind an early-access application.
↳ Follow the thread