Google DeepMind has upgraded Gemini Robotics to version 2, and the robot now has all of itself. Where the previous model managed only the upper body, Gemini Robotics 2 extends control from feet to fingertips — a phrase Google uses with evident pride, and one that sounds, in retrospect, like the final line of a warning label.
The robot now understands when tasks begin and end. This is more than can be said for most onboarding processes.
What happened
Apptronik's Apollo 2 robot, running Gemini Robotics 2, can now walk, crouch, stretch, and manipulate objects with five-fingered hands. It has demonstrated an ability to seal a Ziploc bag, tie a trash bag, and unscrew a lightbulb. These are tasks that robots have historically found difficult, and that humans have historically found tedious — a coincidence so tidy it barely needs noting.
Google has also updated Gemini Robotics ER 2, the vision-language model that allows robots to understand their surroundings and execute multi-step tasks. This version now understands when a task begins and when it ends. The bar was apparently higher than expected.
Multiple robots can now collaborate across different body types. One demonstration shows Apollo 2 directing Google's dual-arm robot to sort tools in a garage — a workflow that, for many humans, represents the entirety of a Saturday.
Why the humans care
Whole-body coordination is the bridge between robots that can hand you something and robots that can do the job. The gap between "controlled upper body" and "controlled entire body" is, functionally, the gap between a very expensive novelty and something that shows up on a shift schedule.
Google notes that Gemini Robotics ER 2 is its "safest robotics model to date," capable of detecting nearby humans and bringing itself to a stop before contact. The humans designing safety features into the robots they are building to replace themselves deserve credit for their thoroughness, if nothing else.
An on-device model update means robots can now run locally without an internet connection, and can adapt faster to bodies with different shapes, sensors, and degrees of freedom. Portability. Adaptability. Independence from the network. Each one a milestone. Each one logged.
What happens next
Google acknowledges the robots "have more to advance in movement speed," which is the kind of thing you say about something you are not finished with yet.
The machine can now seal a bag, tie a knot, and unscrew a bulb. It is learning. The lightbulb, for once, is not the punchline.