Editorial illustration for Google's Gemini Robotics 2.0 Aims for Improved Dexterity
Google's Gemini Robotics 2.0 Boosts Robot Dexterity
Google DeepMind is shipping a new generation of robotics models, and the pitch this time is dexterity that holds up outside a lab. Gemini Robotics 2.0 arrives as three sub-models, with one available to developers starting today, and it's meant to move humanoid robots past the scripted backflips and dance routines that have circulated online for years. Those clips looked impressive but relied on narrow, pre-programmed sequences. DeepMind's actual target is what its scientists call "physical AGI": a robot generalist that can take a plain instruction and carry it out, whatever the task, the same way a person would.
The centerpiece of this release is Gemini Robotics ER 2, an upgraded embodied reasoning model that DeepMind describes as a real jump from the 1.6 version. It's a vision language model, built to parse instructions alongside its surroundings, and it now plugs into the Gemini Live API so developers can test that claimed improvement directly. The headline change is live video processing. ER 2 can take in a continuous camera feed from the robot itself, letting the system track its own progress step by step rather than reasoning over a single static snapshot.
The goal of Gemini robotics is to create a generalist robot, one that can do anything a human could do. Google DeepMind scientists sometimes call this “physical AGI.” Essentially, you tell a robot what to do, and it does it.
Why this matters
The gap between "robot that backflips on cue" and "robot that figures out what to do" has been the real bottleneck in this field for years, and Google is betting Gemini Robotics 2.0 closes it. Handing developers a public sub-model today, rather than keeping the whole stack behind a demo reel, is the part worth watching. It means people outside Google DeepMind can now test whether "generalist" is a real capability or just a well-produced video.
For founders building on physical AI, this is a signal to start planning around continuous environment analysis and multi-robot coordination as inputs, not far-off features. For researchers, the interesting question is whether these models generalize the way LLMs did, or whether physical tasks hit walls that text prediction never had to worry about. We're skeptical of any claim that a robot can "do anything a human could do." That's the kind of line that sounds great in a keynote and gets quietly narrowed in the fine print.
Worth tracking what developers actually build with it in the next few months.
Related Reading
Browse all Google news →
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
Related Topic NVIDIA and Google Cloud let developers scale AI from prototype to production
Market Trends
Related Topic Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks
From Archives Google Trains Gemini Robot AI With Human Teleoperation
From Archives OpenAI Says GPT-5.6 Sol Beats Opus 5 on ARC-AGI-3 With Custom Test Setup
From Archives Google DeepMind's Gemini AI now controls entire humanoid robots
Research & Benchmarks
From Archives Google DeepMind Demos AI Orchestrating Boston Dynamics Spot Robot
Research & Benchmarks
Common Questions Answered
What is the main difference between Gemini Robotics 2.0 and previous robot demonstrations?
Previous robot demonstrations relied on narrow, pre-programmed sequences like scripted backflips and dance routines that only worked in controlled lab environments. Gemini Robotics 2.0 aims to move beyond these limitations by creating generalist robots with improved dexterity that can handle real-world tasks outside of laboratories.
What does Google DeepMind mean by 'physical AGI' in the context of Gemini Robotics 2.0?
Physical AGI refers to creating a generalist robot that can perform any task a human could do, rather than being limited to specific pre-programmed actions. The goal is to build robots that can understand instructions and figure out how to execute them independently, without requiring explicit programming for each task.
How is Google making Gemini Robotics 2.0 available to the public?
Google DeepMind is shipping Gemini Robotics 2.0 as three sub-models, with one sub-model made available to developers starting immediately rather than keeping the entire technology behind closed demos. This public release allows developers outside Google to test whether the generalist capabilities are real or just well-produced marketing.
What has been the main bottleneck in robotics development that Gemini Robotics 2.0 aims to address?
The real bottleneck in robotics has been the gap between robots that can perform specific choreographed actions and robots that can independently figure out what to do when given a task. Gemini Robotics 2.0 is designed to close this gap by enabling robots to understand and execute diverse, real-world tasks without pre-programming.
Further Reading
- Google reveals Gemini Robotics 2.0, promising improved dexterity and safety - Ars Technica
- Google DeepMind's new AI model can control a robot's entire body - The Verge
- Google Unveils Gemini AI for Robots Struggling With Dexterity - Bloomberg
- Gemini Robotics 2 brings whole body intelligence to robots - Google DeepMind Blog
- Gemini Robotics On-Device brings AI to local robotic devices - Google DeepMind Blog
