Google's New Gemini Robotics 2 Can Move Like a Human — But Still Fumbles Simple Tasks

Google DeepMind just gave humanoid robots something most previous models never had: full-body coordination. On July 30, 2026, the company unveiled Gemini Robotics 2, a new AI system that lets robots reason through multi-step tasks while controlling their entire body—from walking and crouching to precise hand manipulation—rather than just directing isolated arm movements the way earlier models did.

But buried in Google's own announcement is an honest admission that undercuts some of the hype: fine motor dexterity, the kind of precise finger control humans do without thinking, remains a genuinely hard problem the model hasn't solved.

Quick Summary & Key Takeaways

• Whole-Body Control, Finally: Unlike the 2025 original, which mainly controlled a robot's upper body, Gemini Robotics 2 can direct an entire humanoid from feet to fingertips.
• Three Models Released Together: Gemini Robotics 2 (motor control), Gemini Robotics ER 2 (multi-step reasoning and planning), and an On-Device 2 version were all released on July 30.
• Dexterity Is Wildly Inconsistent: Success rates for multi-finger tasks range from 32% to 92% depending on the task—strong at simple actions, weak at fine manipulation.
• 92% Success Unscrewing a Light Bulb: But tasks like tying a trash bag or sealing a Ziplock bag had noticeably lower success rates in Google's own tests.
• Early-Access Only: The most capable models are limited to over 100 trusted testers and partners including Apptronik, Agile Robots, and Boston Dynamics, with a developer waitlist now open.

The Three New Models Explained

Model	What It Does	Availability
Gemini Robotics 2	Converts vision and language input into motor commands; controls full humanoids and dexterous hands	Early-access partners and 100+ trusted testers
Gemini Robotics ER 2	The "reasoning brain"—plans multi-step tasks, understands the physical world, coordinates multiple robots	Public preview via AI Studio; private preview on enterprise platform
On-Device 2	A more specialized version designed to run directly on robot hardware	Early-access partners only

What Happened? Inside Google's Robotics Push

According to Carolina Parada, vice president of robotics at Google DeepMind, the company's goal is to bring AI into the physical world and build an intelligence layer usable by every robot—not just Google's own hardware. That framing matters: rather than building robots itself, Google is positioning Gemini as the "brain" that any robotics company can license, similar to how it distributes Gemini across other products.

The technical leap here is whole-body coordination. Earlier robotics AI models were generally good at controlling a single arm or gripper for a narrow task but struggled to coordinate an entire body moving together in real time—walking while reaching, crouching while grabbing an object, balancing while manipulating something with both hands. Gemini Robotics 2 is built specifically to solve that coordination problem, converting camera input and natural-language instructions directly into motor commands across a robot's full frame.

Google was notably candid about the limitations. Kanishka Rao, director of robotics at Google DeepMind, acknowledged that true dexterity remains a distant goal, explaining that robot movements stay slow and deliberate because the machines must consciously pause to think through decisions that humans make intuitively and instantly. He also noted that robots currently learn far less efficiently than people, who can typically adjust their behavior after just one or two mistakes—a gap that remains unsolved even with this upgrade. In testing, the system successfully unscrewed a light bulb 92% of the time, but more complex fine-motor tasks like tying a trash bag or sealing a Ziplock bag showed noticeably lower, less consistent success rates.

Why It Matters: The Long Road Back to Robotics for Google

This release carries weight beyond the technical specs, fitting into a longer, bumpier history:

• A Second Act for Google's Robotics Ambitions: Alphabet acquired a string of robotics startups in the early 2010s, only to wind much of that work down—including shutting its Everyday Robots unit in 2023. Gemini Robotics 2 represents a revived, AI-first approach to a goal Google has chased for over a decade.
• Competition Is Heating Up Fast: Both OpenAI and Nvidia are separately developing their own AI models and software for robots, meaning Google's move is as much about not falling behind as it is about a genuine breakthrough.
• Real Industry Partners Are Already Testing It: Boston Dynamics, Apptronik, and Agile Robots—all serious humanoid and industrial robotics players—are among the early-access partners, suggesting real-world deployment interest beyond a lab demo.
• Honesty About Limits Is Notable: In an industry prone to overstating readiness, Google publishing its own inconsistent dexterity numbers (32% to 92%) is a relatively transparent move that sets more realistic expectations than a polished highlight reel alone would.

💡 AI Tech Safar Insight

The gap between "walks and reasons impressively" and "reliably ties a trash bag" is exactly where the real difficulty in physical AI lives, and Google deserves some credit for not papering over it. Whole-body coordination is a genuinely hard robotics problem, and solving it is a real step forward—but fine dexterity is arguably the harder, more economically important problem, since most real-world labor (cooking, assembly, caregiving) depends on precise hand control far more than walking or crouching. Until that 32%-to-92% range narrows toward consistent reliability, humanoid robots are likely to remain impressive demos and narrow industrial tools rather than the general-purpose household or workplace assistants the industry keeps promising.

Frequently Asked Questions (FAQs)

Q1: What is Gemini Robotics 2?
It's Google DeepMind's newest AI model for robots, released July 30, 2026, that enables whole-body control—allowing humanoid robots to walk, crouch, and manipulate objects while reasoning through multi-step tasks, rather than just controlling isolated arm movements.

Q2: How good is Gemini Robotics 2 at fine dexterity tasks?
It's inconsistent—success rates for multi-finger dexterous tasks range from 32% to 92% depending on the specific task, performing well on simpler actions like unscrewing a light bulb but struggling more with complex manipulation like sealing a Ziplock bag.

Q3: Which companies are testing Gemini Robotics 2?
Early-access partners include Boston Dynamics, Apptronik, and Agile Robots SE, alongside more than 100 trusted testers, with a developer waitlist now open for broader access.

Q4: Is this Google's first robotics AI model?
No. It builds on Gemini Robotics, which debuted in 2025 and primarily controlled a robot's upper body. Gemini Robotics 2 extends that to full whole-body coordination.

What Do You Think?
Is whole-body coordination the bigger breakthrough here, or does fine dexterity remaining stuck at 32-92% success mean humanoid robots are still years away from practical use? Share your thoughts in the comments below!

Related Reading:

• AI Models 2026: The Complete Guide to Gemini, GPT, Claude & China's Challengers
• Claude Opus 5 vs GPT-5.6 Sol: The Benchmark Showdown Developers Are Debating

Source: Reporting based on Bloomberg and Google DeepMind's official blog.