Google DeepMind launches Gemini Robotics 2.5, which doubles the mission success rate compared to previous benchmarks. The Apollo 2 humanoid robot equipped with the model also achieved a 96% success rate in parts sorting tests on Mercedes-Benz production lines. (Previous summary: Google DeepMind released Gemini Robotics 2! Empowering robots with "whole-body intelligence" and moving towards physical AGI) (Background supplement: Google released Gemini 4 Argon! A great evolution of spatial reasoning, driving the Apollo humanoid robot, API pricing released) Google DeepMind announced the launch of Gemini Robotics on November 10 2.5, this vision-language-action (VLA) model allows the robot to convert the environment it sees and the instructions it receives into actual operations. This update focuses on expanding skill range and improving fine movement, on-field reliability and safety. The official post on X lists uses such as folding clothes, delivering cakes, and sorting factory parts. From daily housework to automobile production lines, what Google DeepMind wants to promote is whether robots can learn faster and do things more stably when faced with different objects. Tested on 200 tasks, the highest relative success rate was doubled. Google DeepMind evaluated the new model on 200 different tasks, including holding food with tongs, tying a bow on a gift box, and transferring liquids with a pipette. These actions cover the use of tools, the manipulation of soft objects and the handling of fragile items, and the test is not just about grabbing things. Results announced by the company show that, depending on the type of task, the success rate of Gemini Robotics 2.5 is increased by 30% to 100% compared to the previous test benchmark, up to double, and the speed of completing tasks is also increased by 20% to 30%. Models are available for different robotic hardware, including the Apptronik Apollo 2 humanoid robot with a five-finger manipulator, and the Franka Duo using a two-finger parallel gripper. In other words, this improvement is not just for the display of a single robot. Apollo 2 entered the Mercedes-Benz production line, and the parts sorting test success rate reached 96%. The test is more direct than laboratory operation and comes from Mercedes-Benz's actual operating automobile production line. Google DeepMind collaborated with Apptronik and Mercedes-Benz to conduct a proof-of-concept at the Kecskemét, Hungary, and Sindelfingen plants in Germany, where Apollo 2 handles automotive parts preparation. The robot must take out the metal brackets from the bulk material box, put them into the arranged containers in the order required for each vehicle assembly, and correct any defects in the process. The parts sorting and palletizing trial achieved a 96% success rate. Among them, the embodied reasoning model Gemini Robotics ER is responsible for tracking the workflow, and the Gemini Robotics 2.5 action model controls the hands and arms. This division of labor connects multi-step judgment with actual operations, allowing robots to undertake repeated material preparation, while technicians can leave time for more complex assembly work. Argon uses simulation to accelerate learning, and safety guidance improves along with the movements. Another key to this upgrade is Gemini 4 Argon’s participation in robot research. It can create digital avatars from scenes captured by real cameras, generate thousands of different simulation exercises, and combine them with real operation demonstrations to help the model adapt to new scenes and objects. Argon can also analyze videos of failed operations to help researchers come up with failure reasons and improvement hypotheses, and then test them in simulated environments and real robots. Therefore, training does not have to rely solely on physical machines to repeat trials and errors every time, but still retains real demonstrations and on-site verification. In terms of safety, Gemini Robotics 2.5 adds reasoning guidance when performing actions to avoid physical dangers. The team also launched the open source test benchmark ASIMOV-Jailbreak, which uses unsafe requirements disguised as general tasks to test the jailbreak resistance of robot-related agents. Gemini Robotics 2.5 is becoming available to robotics partners. From repeated practice in a simulated environment to sequential picking and placing of parts on the production line, this update connects learning speed with on-site operations, moving towards practical uses that reduce the burden of repeated material preparation and improve operational stability. Related stories Google shows off 5 Gemini 4 Argon implementations! Create a 3D game with just one tip, and outperform GPT-6 Astra. Din Tai Fung in the United States has begun to introduce robots. The most profitable Chinese restaurant in the United States welcomes tired employees in the backcourt. Robot industry vision: the integration and evolution of automation, artificial intelligence and Web3. The world's first housekeeping robot "Neo" is on sale. It can sweep the floor and make coffee, but does it still need remote human help? 〈Google releases Gemini Robotics 2.5! Driving the Apollo 2 humanoid robot into the Mercedes-Benz production line "Sorting success rate is 96%" This article was first published in BlockTempo "Dongqu Trend - the most influential blockchain news media".