Google DeepMind Unveils Gemini Robotics 2.5
Google DeepMind推出新版机器人AI——Ars Technica
Google DeepMind has released Gemini Robotics 2.5, an updated AI platform for robotics designed to perform physical tasks more accurately, quickly and safely in real-world environments including industrial production settings. The system combines the Gemini Embodied Reasoning model with Gemini Robotics components. According to Google, in lab tests, Gemini Robotics 2.5 showed a 30–100% higher success rate and completed tasks 20–30% faster than the previous version. An Apollo 2 robot equipped with the system achieved a 96% success rate assembling car parts at Mercedes-Benz factories in Europe. The platform is currently available to Google partners such as Agile Robots, Apptronik, and Boston Dynamics.
Source: Chinese robotics — Dataset and collection discovery · Read original article ↗
Article text · Machine translation into English
Google DeepMind launched Gemini Robotics 2.5, which is a new version of its artificial intelligence platform for robotics. As Ars Technica reported, the system is designed to complete real-world physical tasks more accurately, faster, and more safely, including in industrial production environments.
The platform combines the Gemini Embodied Reasoning model and Gemini Robotics components: the former analyzes instructions and camera videos to plan actions, while the latter generates robot motion instructions. DeepMind stated that Gemini Robotics 2.5 performs better in tasks requiring high precision and hand dexterity.
Testing at the Mercedes-Benz factory
During the evaluation period, the model underwent more than 200 task tests, including picking up food with tongs and using a pipette. The company also demonstrated how Apptronik's Apollo 2 humanoid robot ties a bow with five-fingered hands. Manipulating deformable objects, including ribbons, is one of the difficult tasks in robotics technology.
According to Google, in laboratory tests, Gemini Robotics 2.5 had a success rate 30–100% higher than the previous version and completed tasks 20–30% faster. The Apollo 2 robot equipped with this system assembled car parts at several Mercedes-Benz factories in Europe, achieving a success rate of 96%.
Please follow the UA.News Telegram channel for more latest news Telegram.
DeepMind head Kanishka Rao pointed out that it would be very difficult to bridge the remaining 4% gap to achieve near-flawless completion of repetitive work. DeepMind robotics head Karolina Parada emphasized that developers must choose between the highest precision in specific scenarios and the robot's ability to perform a wide range of tasks.
Detecting robot safety
To train this system, DeepMind used Gemini 4 Argon, the latter of which created thousands of simulations based on real camera data. These digital models help generalize new tasks and working conditions faster. Meanwhile, Gemini Robotics 2.5 runs on cloud servers, therefore producing latency when processing instructions.
DeepMind also released the open dataset ASIMOV-Jailbreak, which contains 1,500 prompts used to test whether it is possible to induce robots to perform dangerous behaviors. The company stated that after using Argon in the system, the safety rate against such attempts reached 98.8%, while Gemini Robotics ER 2 achieved 73.5% in Google's benchmarks. Gemini Robotics 2.5 is not yet open to the public, but has been provided to Google's partners, including Agile Robots, Apptronik, and Boston Dynamics.
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
Reported success rate
96%
Reported trials
improvement_range
30–100
View original evidence
据Google称，在实验室测试中，Gemini Robotics 2.5的成功率比上一版本高30–100%，完成任务的速度快20–30%。
Open source S9speed_improvement_range
20–30
View original evidence
据Google称，在实验室测试中，Gemini Robotics 2.5的成功率比上一版本高30–100%，完成任务的速度快20–30%。
Open source S9speed_improvement_range
30–100
View original evidence
据Google称，在实验室测试中，Gemini Robotics 2.5的成功率比上一版本高30–100%，完成任务的速度快20–30%。
Open source S9task_count
>200
Reported trials
- paper: not_reported
What remains unknown
Not established in the collected evidence: Environment, Control, Data origin.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
配备该系统的Apollo 2机器人在梅赛德斯-奔驰位于欧洲的几家工厂装配汽车零部件，成功率达到96%。
Open source S10
该公司表示，在系统中使用Argon后，针对这类尝试的安全率达到98.8%，而Gemini Robotics ER 2在Google基准测试中获得73.5%。
Open source S17
据Google称，在实验室测试中，Gemini Robotics 2.5的成功率比上一版本高30–100%，完成任务的速度快20–30%。
Open source S9
在梅赛德斯-奔驰工厂测试 在评估期间，该模型接受了200多项任务测试，其中包括用夹子夹取食物和使用移液管。
Open source S6
Source：Chinese robotics — Dataset and collection discovery · ua.news