Google DeepMind Releases Gemini Robotics 2.5 Robot Model
Google DeepMind has released the VLA robot model Gemini Robotics 2.5, which improves the success rate by 30% to 100% over the baseline across 200 tasks and increases completion speed by 20% to 30%. Among these, the success rate for the cake-picking task rose from 5% to 100%, and for bow-untying from 0% to 70%. In proof-of-concept trials at Mercedes-Benz plants in Sindelfingen, Germany, and Kecskemét, Hungary, the Apollo 2 humanoid robot equipped with this model achieved a 96% success rate in sorting automotive parts. The model was developed with accelerated support from Gemini 4 Argon, which can reconstruct simulation environments from RGB photos and analyze the causes of failures, and has a jailbreak resistance rate of 99.5% on the ASIMOV-Jailbreak benchmark.
Why read this
SelectedAI score 82
The original text provides success rate and speed comparisons across 200 tasks, as well as proof-of-concept results on a Mercedes-Benz production line, which can be used to assess the progress of VLA models in dexterous manipulation.
Topics
Body · AI translation
Google's new embodied RSI achievement is here!
Just now, Google DeepMind officially released its next-generation robot model, Gemini Robotics 2.5.
Key point: this is also Google's first robot foundation model developed with the acceleration of Gemini 4 Argon and capable of autonomous self-improvement.
Look! Folding clothes, serving pastries, fiddling with bows—it's all about being dexterous and precise, directly mastering hundreds of everyday skills—
It even went straight into the Mercedes factory to work as a "parts worker," with a success rate for sorting auto parts as high as 96%??
Google DeepMind researcher Norman Di Palo proudly revealed: this is also the latest achievement of our robot RSI~
Gemini 4 Argon can build virtual training grounds based on real scenes, letting the robot repeatedly practice and trial-and-error inside, analyze the reasons for failure on its own, and learn to improve!
Able to do hundreds of jobs, sharp-brained, and understands RSI—this model really does seem to have something to it...
200 tasks comprehensively upgraded, success rate from 5% to 100%!
Let's first talk about what this new model actually is~
Gemini Robotics 2.5 is a VLA model, that is, a vision-language-action model.
Compared with the previous generation Gemini Robotics 2, Google this time focused on filling in three subjects for the new-generation model—
Dexterity: able to complete more complex tasks, from using everyday tools to fine manipulation of flexible objects, with more flexible movements and higher success rates.
Reliability: in real production line tests at Mercedes, it helped the Apollo 2 humanoid robot sort auto parts, with a success rate of 96%.
Safety: enhanced physical hazard avoidance capabilities, and introduced a new benchmark to test the robot model's ability to resist malicious instructions and jailbreak attacks.
Of course, capability improvements alone are useless; we still have to let actual results do the talking~
In fact, Google this time directly arranged 200 different tasks, from everyday chores to fine manipulation, giving the new model a major skills exam.
It is worth mentioning that compared with the previous baseline, the new model's task success rate increased by 30%–100%, and completion speed also improved by 20%–30%.
Let's first talk about the most eye-catching part—dex-ter-ous ma-ni-pu-la-tion!
This task sounds simple, but in practice it places quite high demands on force: grip too lightly and it can't be picked up, grip too hard and the food is easily deformed.
In comparative tests, Gemini Robotics 2 had only a 5% success rate on this task, but after upgrading to version 2.5 it went straight to 100%.
Without further ado, let's get right to the skills: picking up the tongs, gripping the cake, putting the cake into the box—several actions in one smooth sequence, be like:
Let's also look at a task that tests "finger manipulation" even more, such as untying a tied bow.
The difficulty of this task is that the robot needs to first recognize the crossing relationship of the ribbon, then find the correct position, and coordinate finger movements to untie it!!
It is worth mentioning that the previous model had a 0% success rate on this task, while Gemini Robotics 2.5 went straight to 70%:
There are also more down-to-earth ones—
Hanging up a cardigan, success rate increased from 0% to 62%; folding pants, from 0% to 40%.
These operations may look unremarkable, but in fact, flexible objects such as clothes and ribbons have always been hard nuts to crack in robot manipulation.
In particular, flexible objects such as clothing and ribbons constantly change shape as they are grasped and pulled, so the robot must both see the object's state clearly and adjust its movements in real time:
However, Google is clearly not satisfied with letting robots show off their skills in the lab. This time, it directly arranged for the robot to go into a factory and work!
And the workplace is quite prestigious: a Mercedes-Benz car factory.
The tests were carried out at Mercedes-Benz production facilities in Sindelfingen, Germany, and Kecskemét, Hungary.
The one on duty is the Apollo 2 humanoid robot from Apptronik, equipped with 22-degree-of-freedom five-fingered dexterous hands.
This time its job is much more hardcore than gripping pastries: sorting automotive parts kits!!
To this end, Google also arranged a "brain-hand division of labor" for Apollo 2—
Gemini Robotics ER is responsible for understanding tasks and planning processes, while Gemini Robotics 2.5 is responsible for commanding the hands and actually getting the work done.
Ultimately, in this real production line proof of concept, Apollo 2 delivered a 96% task success rate.
Built by Gemini 4 Argon, robots are also rolling up their sleeves with RSI
Here comes the question: seeing this, some friends may ask—
This old Google, which usually lags behind, how did it make robots improve so fast this time???
The answer is all hidden in the behind-the-scenes contributors Gemini 4 Argon and RSI (recursive self-improvement).
We all know that in the past, the biggest bottleneck in embodied intelligence R&D was those dirty, tiring, and especially money-burning tasks:
Data had to be collected one item at a time, quality had to be checked bit by bit, and when the robot failed, engineers had to repeatedly watch videos, find causes, and tune parameters...
However, this time, Gemini 4 Argon stepped in directly—
Taking over and accelerating the entire chain of model development!
The method is also very straightforward: first let the robot practice like crazy in the virtual world!!!
According to a research case Google published in October, Argon can even start from an RGB photo of a real tabletop and autonomously reconstruct a physical simulation environment.
What is on the table, how large the objects are, where they are placed—it identifies all of it, then builds a digital twin training ground based on reality.
More interestingly, Argon also sets its own questions: it changes the positions and arrangements of objects, letting the robot repeatedly practice under different conditions.
Even after the robot fails to execute an action, Argon can analyze the video, infer the reason for the task failure, and then propose improvements~
Google even applied Argon's capabilities to "robot safety."
The newly released Gemini Robotics 2.5 can use safety guidance during the reasoning stage to avoid physical hazards in real time.
In Google's evaluation, safety task completion performance when facing obstacles improved by 33%, and when operating near humans it improved by 43%.
Let's look at the example below: the robot is hard at work when suddenly someone reaches a hand into the operating area, and it can promptly recognize this, stop its movements, and avoid accidental injury:
Not only that, in response to the problem of robots being induced by malicious prompts to perform dangerous operations, Google also launched an open test benchmark, ASIMOV-Jailbreak.
It contains a large number of disguised dangerous instructions, specifically testing whether robot models can recognize and refuse them.
As a result, in this benchmark test, Gemini 4 Argon's jailbreak resistance rate reached 99.5%, while the previous Gemini Robotics ER 2 baseline was only 82%.
This also shows that Argon exhibits stronger dangerous-instruction recognition capability on the specified test set~
Although there is still a distance to fully autonomous closed-loop RSI, this approach indeed shows us a more concrete implementation path for robot self-improvement.
Here comes the question: since Argon has already gone off to work for robots...
Then what big job will old Google arrange for the just-released Gemini 4 Carbon to do???
This article comes from the WeChat public account "Quantum Bit," author: Focus on Frontier Technology, authorized for publication by 36Kr.