Google DeepMind's New Robot AI Can Now Control a Full Body, From Feet to Fingertips
Gemini Robotics 2 gives humanoid robots whole-body movement and five-fingered hand control. Here is what the upgrade actually means, and how close these machines are to working alongside people.
Key points
- Google DeepMind released Gemini Robotics 2 on Thursday, upgrading its robot AI to control a humanoid's full body rather than just the upper half.
- The model now drives five-fingered robot hands precise enough to tie a trash bag, seal a zip-lock bag, or unscrew a lightbulb.
- A companion model, Gemini Robotics ER 2, lets robots of different designs work together on shared tasks.
- Google says the safety system can now detect nearby humans and bring a robot to a stop if someone gets too close.
- An on-device version runs without an internet connection and can adapt to robots with very different body shapes.
Until now, Google DeepMind's robot AI told a humanoid machine what to do with its arms and hands. The rest of the body was largely on its own.
Gemini Robotics 2 changes that. Announced Thursday, the new model handles the whole machine, from how it plants its feet to how it curls its fingers. Google calls this "whole-body coordination," meaning a robot can now crouch, walk, reach and grip in one coordinated sequence rather than several disjointed ones.
What can the robots actually do now?
More than before, but not everything yet. Demo videos show Apptronik's Apollo 2 robot bending down to pick up a watering can and scanning a shelf to find a specific item.
The hand work is the detail worth paying attention to. Earlier versions controlled simpler gripper-style hands. Gemini Robotics 2 supports full five-fingered hands, which opens up tasks that need a pinching or twisting motion: sealing a zip-lock bag, knotting a trash bag, removing a light bulb from its socket. These are the small manual tasks that have historically defeated robots, and we tracked the first signs of this capability in our 30 July story on Gemini Robotics 2.
Google is honest that speed's still a weak point. The company says its robots "have more to advance in movement speed," which is a polite way of saying they're still slow by human standards.
What is the ER model, and why does it matter?
Gemini Robotics ER 2 is the "embodied reasoning" layer, a vision-language model (software that reads both images and text) that helps robots analyze their surroundings and carry out multi-step tasks over a longer stretch of time.
The core upgrade is task awareness. Earlier, the system could get confused about where one job ended and another began. Google says ER 2 "now understands when tasks begin and end," which sounds minor but matters enormously in a warehouse or kitchen where jobs blur together.
ER 2 also enables different robot types to collaborate. One video shows Apollo 2 directing Google's dual-arm robot to sort tools into a bin. Two machines, different designs, one shared job.
Should people working near robots be worried?
Google says Gemini Robotics ER 2 is its safest robotics model yet. It can detect when a person gets too close and stop the robot before contact occurs.
That's a baseline safety feature, not a complete answer to workplace safety questions. But it's a meaningful step as these machines start moving their whole bodies in shared spaces.
The on-device model, which runs locally without any internet connection, can now adapt faster when installed in a robot with a different body shape or sensor setup. Manufacturers won't need to rebuild from scratch every time hardware changes.
Dexterity is closing faster than most people expected. Repetitive manual work in structured environments is the near-term target, not skilled trades. Watch the speed gap next: that's the remaining barrier between demonstration videos and a robot actually keeping pace with a person on a shift.
