Google Launches Gemini Robotics 2, Full-Body Coordination Still Faces Dexterity Bottleneck
Taylor Wilson
Google DeepMind released Gemini Robotics 2, extending AI control from a robot's upper body to full-body coordination for the first time; yet the team concedes true dexterity remains a distant goal — the robot still has to stop and think through motions humans do by instinct, capping the pace of commercialization.
What exactly got upgraded?
The core leap: control scope expands from the upper body to the entire body — the robot can now walk, crouch, and manipulate objects at the same time.
This means → the robot is no longer an "arm bolted to a desk." It can move and handle things simultaneously.
In plain terms = it went from sitting at the workbench to walking around the room — simple-sounding, but a fundamentally harder control problem for AI.
What do the three models each do?
Gemini Robotics 2 converts camera input and natural-language commands into motor actions — the robot's "cerebellum."
Gemini Robotics ER 2 handles high-level task planning, multi-step coordination, and multi-robot collaboration — the "dispatch center."
On-Device 2 runs locally on the hardware, no cloud required. All three can work together or independently.
Why is dexterity still the bottleneck?
Google DeepMind robotics director Kanishka Rao conceded that true dexterity remains a distant goal; movements are still slow and deliberate.
This reflects a fundamental gap — motions humans complete by instinct require the robot to pause and compute step by step.
Learning efficiency compounds the problem: humans adjust after one or two mistakes; current systems cannot match that.
In plain terms = the system screws in a light bulb at 92% success, but tasks like tying a trash bag or sealing a zip-lock bag score significantly lower.
What changed on safety?
Google calls ER 2 its safest robot model to date, with improvements in following instructions and avoiding people.
A new benchmark tests whether the robot can recognize uncertainty and actively refuse unsafe commands.
This means → safety is not just "don't bump into anyone" — the robot must also judge "I should not execute this instruction."
Who gets access, and how?
ER 2 is available through Google's AI Studio; enterprise access enters private preview. Robotics 2 and On-Device 2 open to early partners and over 100 trusted testers.
Key hardware partners include Apptronik, Agile Robots, and Boston Dynamics.
This reflects Google's playbook: build the AI brain in-house, outsource the body to hardware partners — setting up a three-way race with OpenAI and Nvidia in embodied intelligence.
Who will commercialize embodied AI first?
Three diverging approaches: Google builds a general control model; OpenAI pursues a foundation model fusing vision, language, and action; Nvidia supplies the training toolchain.
All three face the same bottleneck: dexterity. Whichever cracks human-like finger coordination first opens the door to commercial deployment.
In plain terms = everyone can make a robot walk and carry a box; until it can tie a trash bag, it stays out of real homes and factories.
Content is for reference only, not financial advice.