I'm going to tell you something.The shadows
The robot forward reported on November 11, last night, that the man was killed in a plane crash.Google DeepMind has officially released a new generation of VLA model Gemini Robotics 2.5.According to Google DeepMind, this is its most powerful VLA model to date, not only expanding the robot's operational capabilities, but also further improving the success rate and execution speed. Compared to the previous model, the model improved the success rate in some tasks by up to 100%, with execution speeds up to 30%, and achieved the task success rate of 96% in the Mercedes-Benz automotive line.
However, more attention is paid to its reliability in real-life scenarios. Some netizens say that even if the test shows the accuracy of picking up parts to 96%, the remaining 4% of errors on the assembly line are critical.
01.
More than 200 missions have been tested.The success rate of the Mercedes-Benz production line is 96%.Compared to previous models in the Gemini Robotics series, Gemini Robotics 2.5 is primarily based on theFlexibility, reliability and safetyIn the real world, robots not only need to perform simple crawling and placement, but also to cope with a variety of complex operations, including using tools, handling deformable flexible objects, and handling detachable objects. For this reason, Google DeepMind tested the operational capabilities of Gemini Robotics 2.5 for more than 200 different tasks.Gemini Robotics 2.5 improved relative success rates from 30% to 100% in different types of missions, and the speed of mission execution from 20% to 30%.In particular, the performance of Gemini Robotics 2.5 has improved significantly in multi-step and contact-intensive tasks. For example, the success rate of the model has improved in the task of closing the black box seals.Increased from 0% to 93%Take the hanging cups off and put them on the table;Increased from 29% to 75%• Successful cleaning of garbage using brushes;Increased from 0% to 52%。
For more sophisticated and sophisticated operations, the model has also made significant progress.Increased from 5% to 100%The success rate of solving the ribbon butterfly knotIncreased from 0% to 70%Success rates of hanging from a shirt;From 0% to 62%The success rate of folding the knot isIncreased from 0% to 40%。
In addition, Gemini Robotics 2.5 still supports a variety of different types of robot cores, including the sophisticated Apollo 2 with SharpaWave and the Franka Duo with clawed claws. In addition to operating in a laboratory environment, Google DeepMind has also partnered with Apptronik, Mercedes to bring the model to the automotive production line. The PoC's mission is to sort and sort automotive parts.In the end, Apollo 2 achieved a 96% success rate in the mission.According to Jeff Cardenas, co-founder and CEO of Apptronik, the performance of the model improved significantly after just a few weeks of training, and the performance decline was small after moving to the actual production line of the Mercedes.02.
Improved safety control capabilities,Gemini 4 Argon Accelerated TrainingIn addition to its sophistication and reliability, Gemini Robotics 2.5 further enhances the ability of robots to operate safely in complex environments. In Gemini Robotics 2.5 Google DeepMind introduced a reasoning time guideline mechanism that allows robots to adjust their movements in real time during tasks to avoid physical hazards around them.The mechanism improved the success rate of robots performing tasks safely near obstacles by 33% and 43% near people.According to Google DeepMind,Gemini 4 Argon played an important role in the development of Gemini Robotics 2.5.In addition, the Gemini 4 Argon can serve as a simulation engine based on real-life scenarios, create realistic digital environments, and generate thousands of different simulation exercises. Combined with real-life training demonstrations, these simulations can help the robot adapt more quickly to new situations and objects. In addition, the Gemini 4 Argon can predict the reasons for mission failure based on the robot's video, and test improvements on the simulation environment and real robots.
In addition to accelerated robot training, Gemini 4 Argon has also performed well in security assessments. Google DeepMind has opened up the security assessment benchmark ASIMOV-Jailbreak, which contains hundreds of test cues covering dozens of scenarios, mainly for testing the model's ability to resist disguised hazard instructions and jailbreak attacks.The Gemini 4 Argon has a jailbreak resistance of 99.5%, up from 82% of the previous Gemini Robotics ER 2.。03.
I'm not going to say that I'm not going to be able to do it.Real operations are still key.This upgrade to Gemini Robotics 2.5 not only reflects the advances in VLA models in robotics operations, but also shows more possibilities for general large models in robotics training and development. Using large models to generate simulation data, analyze the causes of mission failures, improve the efficiency of robot learning of new skills, and reduce reliance on real machine data. However, continuous improvement in mission success rate, reliability, and security remains key to the smart move towards scaling down.
