DeepMind adds whole-body control to Gemini Robotics 2

DeepMind extended Gemini Robotics from upper-body-only control to full-body locomotion on July 30. The update lets a humanoid walk, crouch, reach and manipulate as one motion, per the company's blog post. That's the difference between a robot that can grab something handed to it and one that can walk across a room to get it. Benchmarks in the same post show that gap isn't closed: general whole-body manipulation across shelf, table and floor tasks scored 45.7 to 76.3 percent.

The release comes in three pieces. Gemini Robotics 2 VLA maps vision and language to motor commands "from feet to fingertips." ER 2 is an embodied-reasoning model built for multi-step planning, and On-Device 2 is the smaller version meant to run locally.

Testing covered five robot platforms: Apptronik's Apollo 2, Franka's Duo, Dexmate, SO101 and Trossen, according to the post.

Gripper pick-place and insertion tasks scored 74.2 to 89.6 percent, a tighter band that suggests two-finger grasping is close to reliable. Multi-finger manipulation tells a different story: 32 to 92 percent, a 60-point spread the post doesn't break down by task or hand design. The announcement drew 606 points on Hacker News.

For builders, the walking demo is the easy sell. The real number is the 32-to-92 spread on multi-finger tasks, still the least predictable part of the system. Closing that range is what separates a robot that works in a demo video from one that works in a warehouse.