IT Home On November 10, Google introduced the most powerful Visual-Language-Action (VLA) robot model to date: Gemini Robotics 2.5, which significantly increases the range of skills a robot can learn and its learning speed.
The model enhances the dexterity, reliability, and safety of robots, enabling them to leave the laboratory and enter more unstructured real environments, thereby assisting humans in completing a wider range of tasks.
Earlier this year, Google introduced whole-body intelligence for various robots, including humanoid robots. This progress was largely thanks to the ability of the Gemini 4 Argon acceleration robot development, which further expanded the boundaries of physical intelligence.
Google tested the task scope, performance, and complexity of Gemini Robotics 2.5 in over 200 different tasks.
Gemini Robotics 2.5 has achieved a leap in three key areas:
Agility: Robots can use everyday tools to handle delicate soft objects and perform significantly more tasks, with improved performance and agility.
Reliability: In a real-world proof-of-concept on Mercedes-Benz production lines, this model helped Apptronik’s Apollo 2 humanoid robot pick and pack automotive components, achieving a success rate of 96%. It can handle monotonous tasks, allowing technicians to focus on complex assembly tasks.
Security: As robots enter daily environments, Google has improved physical security guidance and introduced new testing benchmarks for robots against jailbreak attempts.
Currently, Gemini Robotics 2.5 is being made available to Google’s robot partners.
The real environment is unstructured and unpredictable, so robots must have both high precision and high reliability to handle everyday objects. With Gemini Robotics 2.5, Google has expanded the capabilities of the model in terms of task breadth, performance, and complexity. 200 tests have been conducted, such as distributing food with clips, tying bowties on gift boxes, and transferring liquids.
These tasks integrate advanced tool usage, the manipulation of complex deformable objects, and delicate handling of fragile items—representing a step toward human-level dexterity.
The evaluation shows that the model has significant improvements in task accuracy, robustness, and speed: compared to the previous baseline, Gemini Robotics 2.5 has a success rate of 30% to 100% higher depending on the task type, and task completion time is 20% to 30% faster.
Same as the previous Gemini Robotics model, it is natively compatible with various robot types, including the Apptronik Apollo 2 humanoid robot with a SharpaWave hand having five fingers and 22 degrees of freedom, as well as the Franka Duo platform equipped with two parallel grippers.
The ultimate test for robot models is whether they can perform tasks smoothly in real-world working environments. Google, in collaboration with Apptronik and Mercedes-Benz, has deployed the Apollo 2 humanoid robot to test Gemini Robotics 2.5 on automobile production lines in Kecskemét, Hungary, and Sindelfingen, Germany.
Robots find it difficult to adapt to completely new, unpredictable environments, but the Gemini Robotics model gives Apollo the ability to think, perform multi-step reasoning, and adjust its actions as needed.
In this concept validation, Apollo 2 autonomously completed the vehicle parts picking and sorting task: This is a highly repetitive and physically demanding task that requires sorting individual components from loose boxes into organized pallets according to the precise requirements of each vehicle online. IT Home has provided the relevant video below:
By having robots handle this monotonous preparation work, the production line can operate smoothly, and technical staff can free up their energy to focus on complex assembly tasks.
Google uses the embodied reasoning model Gemini Robotics ER to track workflows, with the Gemini Robotics 2.5 action model controlling the robot’s arms and hands: the robot inserts into the feeding box, picks up the metal car bracket, places it into the sorting container, and continuously adjusts the assembly gap.
The pilot test success rate reached 96%, showing a promising future for its application in large-scale industrial environments. When robots operate in messy daily spaces, physical safety is the top concern. Google has significantly upgraded the safety features of Gemini Robotics 2.5.
Gemini Robotics 2.5 uses a reasoning guidance mechanism during reasoning, which can avoid physical dangers in real time. Just as the safety classifier adds a protective layer to Gemini’s response, this guidance capability enables the robot to safely navigate through complex environments.
After evaluation, this capability increases the completion rate of safe tasks around obstacles for robots by 33%, and increases the completion rate of safe tasks around people by 43%.
Google is also advancing the security capabilities of long-term intelligent agent tasks: Gemini 4 Argon is at the leading level in agent robot evaluations, with evaluation dimensions including rejection of dangerous tasks, uncertainty resolution, and close human-robot collaboration.
To test the robustness of jailbreak resistance, Google has released an open-source benchmark ASIMOV-Jailbreak, which includes hundreds of prompts disguised as dangerous requests to cover dozens of scenarios. The jailbreak resistance rate of Gemini 4 Argon is 99.5%, a significant improvement over the 82% of the previous baseline Gemini Robotics ER 2.
To teach robots new skills, researchers need massive physical data and efficient computing methods. The robot development capability of Gemini 4 Argon can serve as a simulation engine to overcome these bottlenecks and significantly accelerate the development of Gemini Robotics 2.5.
Argon can generate thousands of diverse simulated practice segments based on real camera scenes to build high-fidelity digital twins. By combining real training demonstration data, robots can adapt more quickly to new scenarios and objects.
Argon can also help researchers infer the reasons for task failures from video clips, and conduct synchronized testing of optimal improvement solutions on real robots in simulation environments.
With the acceleration provided by cutting-edge intelligent models like Gemini 4 Argon, Gemini Robotics 2.5 has become an important milestone on the path toward universal physical intelligence. Google continues to improve in three key areas: dexterity, real-world reliability, and strict safety, in order to create physical intelligence that enables machines and humans to jointly solve real-world challenges.