- Gemini Robotics 2.5 achieved a 96% success rate in factory tests.
- The model can translate visual information into physical actions for robots.
- Google has improved the model's ability to handle physical hazards.
- Gemini 4 Argon has accelerated the development speed of robotic models.
Google DeepMind has announced the launch of Gemini Robotics 2.5, a model designed to help robots learn physical tasks more quickly, handle delicate items, and operate more reliably in real-world environments. The model achieved a 96% success rate in a factory proof-of-concept involving Apptronik's Apollo 2 humanoid robot, conducted on the Mercedes-Benz production line. These robots automatically sort car parts from bulk containers into organized trays and prepare components in the required order for vehicle assembly.
The trials were conducted at production facilities in Kecskemét, Hungary, and Sindelfingen, Germany. Google stated that this work could help automate repetitive and physically demanding tasks while allowing factory technicians to focus on more complex assembly work.
Features and Applications of Gemini 2.5
Gemini Robotics 2.5 is the company's latest visual-language-action (VLA) model, capable of translating visual information and commands into physical actions for robots. The model is currently being rolled out to Google's robotic partners.
The model has been evaluated on over 200 different tasks, including serving food with tongs, tying bows on gift boxes, and transferring liquids using pipettes. These activities tested a variety of capabilities, from handling delicate items to manipulating flexible materials and using everyday tools. Compared to previous benchmarks, the model improved task success rates by 30% to 100%, and according to Google, its task completion speed also increased by 20% to 30%. The model can operate on various robotic platforms, including Apptronik's Apollo 2 humanoid robot and the Franka Duo system, which uses parallel grippers.
In a demonstration at Mercedes-Benz, Google's embodiment reasoning model Gemini Robotics ER tracked workflows, while Gemini Robotics 2.5 controlled the robot's arms and hands. Apollo 2 retrieved metal brackets from supply bins and placed them into sorting containers, adjusting its movements as necessary.
Enhanced Safety and Adaptability
In addition to flexibility, Google has improved the model's ability to handle physical hazards. Its reasoning time guidance helps robots avoid obstacles and operate among people. In the company's evaluations, this capability increased the success rate of safety tasks around obstacles by 33% and by 43% when near personnel. Google also introduced ASIMOV-Jailbreak, an open-source benchmark for testing whether robotic systems can resist disguised hazardous commands. This benchmark includes hundreds of prompts covering dozens of scenarios.
Google reported that Gemini 4 Argon achieved a 99.5% resistance rate in these jailbreak attempts, compared to 82% for the previous Gemini Robotics ER 2 benchmark. The company also stated that Gemini 4 Argon accelerated the development of its new robotic models. The system can generate thousands of simulated practice sessions from real-world camera footage and create digital twins of physical environments. Researchers can combine these simulations with actual training demonstrations to help robots generalize their skills to unfamiliar objects and environments.
Argon can also analyze failed attempts and provide improvement suggestions for subsequent training.
These advancements mark Google's latest push to introduce more adaptable robots into industrial and everyday environments. However, the 96% factory results remain a proof of concept rather than evidence of widespread commercial deployment. Gemini Robotics 2.5 is currently being rolled out to Google's robotic partners.
Commercial Potential and Challenges of Gemini Robotics 2.5
The launch of Gemini Robotics 2.5 demonstrates Google's significant progress in robotics technology, particularly in factory automation. While the 96% success rate in testing is encouraging, it is important to note that this is merely a proof of concept, and future commercial applications remain uncertain. Additionally, improvements in handling physical hazards and enhancing safety indicate its potential in everyday environments. As the technology continues to develop, effectively commercializing these robotic systems will be a challenge that Google needs to address.