Gemini Robotics 2 controls a humanoid's body from feet to fingertips

Google DeepMind has announced a new version of its artificial intelligence model for robotics — Gemini Robotics 2. Whereas the previous version was responsible for controlling only the upper body of a humanoid, the system now coordinates the movements of the whole body — “from feet to fingertips.”

Full-body control

With support for whole-body motion, robots gain a broader repertoire of actions: they can walk, crouch, bend down and manipulate objects. In the videos Google released, Apptronik's Apollo 2 robot bends down to pick up a watering can, and also locates and takes specific items off a shelf.

DeepMind acknowledges that robots still have room to grow when it comes to speed of movement, but calls the update an important step toward the skills needed to perform complex real-world tasks that require the coordinated work of the entire body.

Fine motor control has been improved separately: the model now operates more sophisticated five-fingered hands. This makes it possible to seal a zip-lock bag, tie off a garbage bag or unscrew a light bulb.

A reasoning model and teamwork

Alongside this, Gemini Robotics ER (embodied reasoning) has been updated — a model at the intersection of vision and language that helps robots analyze their surroundings, process instructions and carry out multi-step tasks. The ER 2 version handles long-running tasks better and “understands when a task begins and when it ends.”

The new model also lets several robots of different types work together. In one video, Apollo 2 directs Google's two-armed robot, tasking it with placing tools into a bin while cleaning up a garage.

DeepMind calls Gemini Robotics ER 2 its “safest robotics model to date”: it more reliably detects the presence of people nearby, triggers protective mechanisms and stops the robot if someone comes too close.

The company has also refined the Gemini Robotics On-Device Model, which runs locally on the robot itself without an internet connection. It now adapts more quickly to new “bodies,” including devices with a fundamentally different shape, set of sensors and number of degrees of freedom.

Source: The Verge