Google Unveils Gemini Robotics 2.5, a VLA Model with Enhanced Dexterity and Reliability

IT Home On November 10, Google released the most capable Vision-Language-Action (VLA) robot model to date Gemini Robotics 2.5, significantly improving the breadth of skills and learning speed that robots can acquire.

The model enhances the dexterity, reliability, and safety of robots, helping robots leave the laboratory, enter more unstructured real-world environments, and assist humans in completing more types of tasks.

Earlier this year, Google already introduced it for various types of robots, including humanoid robots Full-body intelligence.This progress was largely benefited by Gemini 4 Argon's ability to accelerate robot R&D, further expanding the boundaries of physical intelligence.

Google in over 200 items The task breadth, performance, and complexity of Gemini Robotics 2.5 were tested across different tasks.

Gemini Robotics 2.5 achieved leaps in three key areas:

• Dexterity: Robots can use everyday tools, handle delicate flexible objects, complete significantly more tasks, and performance and dexterity are greatly improved.

• reliability: In a real-world concept validation on a Mercedes production line, this model assisted Apptronik's Apollo 2 humanoid robot in picking and packing car parts, with a success rate of 96%, capable of handling monotonous work, allowing technicians to focus on complex assembly tasks.

• Safety: As robots enter everyday environments, Google optimized physical safety steerability and launched a brand-new testing benchmark for robot adversarial jailbreaking.

Currently, Gemini Robotics 2.5 is open to Google's robotics partners.

Real-world environments are unstructured and unpredictable, and robots must possess both high precision and high reliability to handle everyday objects. With Gemini Robotics 2.5, Google has expanded the model's capabilities in task breadth, performance, and complexity, with test tasks reaching 200 items, For example, distributing food with tongs, tying bows on gift boxes, and pipetting liquids, etc.

These tasks integrate multiple skills such as advanced tool use, complex deformable object manipulation, and delicate handling of fragile items, representing a step toward human-level dexterity.

Evaluations show that the model has significant improvements in task accuracy, robustness, and speed: compared to the previous baseline, Gemini Robotics 2.5, depending on the task type, has a success rate higher by 30% to 100%, and task completion speed faster by 20% to 30%.

Consistent with the previous generation Gemini Robotics model, it natively adapts to multiple robot types, including the Apptronik Apollo 2 humanoid robot equipped with a five-fingered 22-degree-of-freedom SharpaWave hand, as well as the Franka Duo platform equipped with a two-fingered parallel gripper.

The ultimate test for a robot model is whether it can stably execute tasks in real work scenarios. Google, in collaboration with Apptronik and Mercedes-Benz, deployed Apollo 2 humanoid robots to test Gemini Robotics 2.5 on active automobile production lines in Kecskemét, Hungary, and Sindelfingen, Germany.

Robots find it difficult to adapt to entirely new unpredictable environments, while the Gemini Robotics model gives Apollo the ability to think, perform multi-step reasoning, and adjust actions according to changes.

In this proof of concept, Apollo 2 autonomously completed the automotive parts picking task: this is a highly repetitive, physically demanding task that requires sorting individual components from bulk bins into organized pallets according to the precise requirements of each vehicle on the line. IT Home attaches the relevant video below:

Entrusting such monotonous preparation work to robots can both ensure smooth assembly line operation and free up technical personnel 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 reaches into the supply bin, picks up a metal car bracket, places it into a sorted container, and corrects assembly gaps throughout the process.

The pilot test success rate reached 96%, demonstrating considerable prospects for large-scale industrial scenario landing applications. When robots operate in cluttered daily spaces, physical safety is the primary consideration. Google significantly upgraded the safety features of Gemini Robotics 2.5.

Gemini Robotics 2.5 adopts an inference-time guidance mechanism, capable of avoiding physical hazards in real-time. Just as safety classifiers add a protective layer to Gemini's responses, this guidance capability allows robots to navigate safely through complex environments.

Upon evaluation, this capability improved the completion rate of safe tasks for robots around obstacles 33%, the completion rate of safety tasks around personnel is improved 43%.

Google is simultaneously advancing the safety capabilities of long-cycle agent tasks: Gemini 4 Argon is at a leading level in agent robot evaluations, with evaluation dimensions including refusal of dangerous tasks, uncertainty resolution, and close-range human-robot collaboration.

To test jailbreak robustness, Google released an open-source benchmark ASIMOV-Jailbreak, containing hundreds of prompts disguised as dangerous requests, covering dozens of categories of scenarios. Gemini 4 Argon's jailbreak resistance rate reaches 99.5%, a significant improvement over the 82% of the previous generation baseline Gemini Robotics ER 2.

To teach robots to master new skills, researchers need massive amounts of physical data and efficient construction algorithms. Gemini 4 Argon's robot R&D capabilities can serve as a simulation engine, solving the above bottlenecks and significantly accelerating the development progress 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. Combined with real training demonstration data, robots can generalize and adapt to entirely new scenes and objects faster.

Argon can also help researchers deduce the reasons for task failures from video clips and synchronously test the optimal improvement plans in simulation environments and on real robots.

Accelerated by cutting-edge intelligent models like Gemini 4 Argon, Gemini Robotics 2.5 becomes an important milestone on the road toward physical general intelligence. Google continues to refine three dimensions—dexterity, real-world reliability, and rigorous safety—to create the physical intelligence required for robots and humans to collaborate in solving real-world challenges.