Google DeepMind's Gemini Robotics 2.5 significantly improves the physical dexterity of robots, increasing success rates by up to 100% and reducing task completion times by up to 30%, while achieving a 96% success rate in real-world factory settings.
Imagine this: On a factory workbench cluttered with complex parts, a robotic arm effortlessly picks out the exact components needed without hesitation, and performs intricate knot-tying tasks—previously done by human hands—in the blink of an eye. For a long time, robots were little more than “simple machines” that repeated pre-set paths, but they are now evolving into “smart workers” that can analyze situations and act on their own.
On November 10, 2026, Google DeepMind unveiled “Gemini Robotics 2.5,” which elevates robot intelligence to the next level. It boasts the highest performance among all Vision-Language-Action (VLA) models Google has introduced to date [Reference 9, Reference 12].
Why Does This Matter?
The ability for robots to move beyond simple repetitive tasks to understand situations and demonstrate “dexterity” signals massive changes for our industrial sites and daily lives. Past robots were often specialized for specific processes, requiring human intervention to rewrite code whenever the situation changed slightly [Reference 5]. Now, however, robots can perceive their environment and make judgments themselves, enabling them to perform much more general-purpose roles.
In particular, if robots can perform precise tasks in sorting centers or complex assembly lines as well as, or better than, humans, productivity will improve dramatically.
Simplified: A Robot with ‘Eyes’ and a ‘Brain’
To put it simply, Gemini Robotics 2.5 is like giving a robot “situational awareness” and “flexible fingers.” If existing robots were dolls moving blindly according to set commands, robots equipped with this model look through cameras (eyes) to see their surroundings, use AI (brain) to decide what needs to be done within those images, and then move their robotic arms (hands) to execute immediate actions [Reference 9].
The core VLA model plays the role of connecting “Vision” information with “Language” instructions to generate robot “Actions.” It is very similar to the process of reading a cookbook (language), checking the state of ingredients with our eyes (vision), and then chopping them (action). Thanks to this, robots can find and perform tasks on their own even in complex environments [Reference 10].
Current Status: Performance Proven by Numbers
Gemini Robotics 2.5 is already proving its capabilities in real-world settings. According to Google DeepMind, this model has improved task success rates by 30% to up to 100%, depending on the task, compared to previous generations of robot models [Reference 12].
The speed is also remarkable. It is not just the success rate that has improved; task completion time has also become 20% to 30% faster than before [Reference 12]. In a pilot operation conducted at a Mercedes-Benz factory, the Apptronik “Apollo 2” humanoid robot achieved an impressive 96% success rate using Gemini Robotics 2.5 [Reference 11, Reference 12].
What Happens Next?
Google aims to use this technology to make robots more general-purpose and useful in the physical world. Since the Gemini Robotics 2.0 model, it has become possible to handle full-body control of humanoid robots and even coordinate multiple robots working together in the same space [Reference 10].
Of course, this does not mean that “Physical Artificial General Intelligence (AGI)” has arrived immediately. Safety issues and technical challenges still remain to be solved [Reference 5, Reference 12]. However, judging by the pace of technological development, the future where robots safely and precisely perform human work right beside us is no longer a distant story.
AI Opinion
Gemini Robotics 2.5 has proven the potential of “actionable intelligence” that does more than just think—it gets actual work done in the physical world. The day we see robot colleagues silently helping us by our side is now much closer.
References
- Gemini Robotics 2.5 accelerates real-world general dexterity
-
[Gemini— Google DeepMind Our next era of frontier intelligence](https://deepmind.google/models/gemini/) - Google Just Unveiled Gemini Robotics, and It’s a Whole New Level of…
-
[Google DeepMind’s Gemini Robotics 2 boosts AI accuracy… LinkedIn](https://www.linkedin.com/posts/aslam-ahamed_artificialintelligence-ai-robotics-activity-7490016683873542144-3VmV) -
[Google reveals Gemini Robotics 2.0, promising improved dexterity and safety AIBlog.com.my](https://aiblog.com.my/articles/google-reveals-gemini-robotics-20-promising-improved-dexterity-and-safety) - Gemini Robotics 2 advances dexterity
- Google DeepMind releases Gemini Robotics 2.5 for robot dexterity
- Gemini Robotics — Google DeepMind
- Gemini Robotics 2.5 Enhances Dexterity and Safety in Real …
- Gemini 2.5 for robotics and embodied intelligence - Google …
- Improved task success rates and reduced speed
- 2x battery life increase
- New robot hardware design
- Analyzing past data logs only
- Camera inputs and instructions
- User voice commands only
- 85%
- 90%
- 96%