AI News

Google DeepMind Releases Gemini Robotics 2.5, Achieving a 96% Success Rate in Real-World Testing on Mercedes Production Lines

ZhidongxiUpdated Nov 12 03:58

AI summary

Google DeepMind has released its next-generation VLA model, Gemini Robotics 2.5, claiming that across tests on more than 200 tasks, its success rate improved by a relative 30% to 100%, and its execution speed increased by 20% to 30%. In an automotive parts sorting PoC in collaboration with Apptronik and Mercedes, the Apollo 2 equipped with this model achieved a 96% task success rate. The model also introduces an inference-time guidance mechanism, improving the success rates of safety-related tasks near obstacles and near people by 33% and 43%, respectively.

Why read this

AI score 71

The original text lists the item-by-item success rate changes across more than 200 tasks and the 96% real-world test result on the Mercedes production line, which can be used to assess the boundaries of progress for VLA models in real-world manipulation.

Topics

Tags

Body · AI translation

Robot Frontier reported on November 11 that last night, Google DeepMind officially released its new-generation VLA model Gemini Robotics 2.5, saying it can learn and complete over a hundred everyday operations, such as folding clothes, delivering snacks, and sorting auto parts on factory production lines.

Google DeepMind said this is its most capable VLA model to date, not only expanding the robot's manipulation capabilities but also further improving task success rates and execution speed.

Compared with previous models, this model improves success rates by up to 100% on some tasks and boosts execution speed by up to 30%, and achieved a 96% task success rate on the Mercedes-Benz automotive production line.

Some netizens joked in the comments, "It can fold clothes and also work on the Mercedes-Benz production line—turns out robots have to work two jobs too."

However, most people focused on its reliability in real-world scenarios. One netizen said, "Even if tests show a 96% accuracy rate for picking parts, on an assembly line, the remaining 4% of failures are what really matter."

1. Tested on over 200 tasks, 96% success rate on the Mercedes-Benz production line

Compared with previous Gemini Robotics series models, Gemini Robotics 2.5 mainly achieves improvements in three areas: dexterity, reliability, and safety.

Among these, dexterity is a major focus of this upgrade.

In real environments, robots not only need to complete simple grasping and placing, but also need to handle various complex operations, including using tools, handling easily deformable flexible objects, and finely manipulating fragile items.

To this end, Google DeepMind tested the manipulation capabilities of Gemini Robotics 2.5 on more than 200 different tasks. Test results show that, compared with the previous baseline model, Gemini Robotics 2.5 achieved relative success rate improvements of 30% to 100% across different types of tasks, and improved task execution speed by 20% to 30%.

Specifically, Gemini Robotics 2.5 showed significant improvement in multi-step and contact-intensive tasks.

For example, in the task of closing the lid of a black toolbox, the model's success rate rose from 0% to 93%; the success rate of taking down a hanging cup and placing it on a tabletop rose from 29% to 75%; the success rate of using a brush to sweep trash into a dustpan rose from 0% to 52%.

For more complex dexterous manipulation, the model also made notable progress.

In the task of using tongs to pick up pastries, the success rate rose from 5% to 100%; the success rate of untying a ribbon bow rose from 0% to 70%; the success rate of hanging a cardigan rose from 0% to 62%; and the success rate of folding pants rose from 0% to 40%.

In addition, Gemini Robotics 2.5 still supports a variety of different robot embodiments, including Apollo 2 equipped with the SharpaWave dexterous hand and the Franka Duo with a gripper.

Beyond manipulation capabilities in laboratory environments, Google DeepMind also collaborated with Apptronik and Mercedes-Benz to bring the model to an automotive production line.

The task of this PoC was automotive parts sorting and sequencing. Apollo 2 needed to take designated parts from bins containing large numbers of parts and place them into corresponding trays according to the assembly needs of vehicles on the production line.

In this process, Gemini Robotics ER was responsible for tracking the entire task flow, while Gemini Robotics 2.5 was responsible for controlling the robot's hands and arms. Ultimately, Apollo 2 achieved a 96% success rate on this task.

Apptronik co-founder and CEO Jeff Cardenas said that after just a few weeks of post-training, the model's performance improved significantly, and after transferring to Mercedes-Benz's real production line, the performance degradation was very small.

2. Improved safe control capabilities, Gemini 4 Argon accelerates training

In addition to dexterity and reliability, Gemini Robotics 2.5 further improves the robot's safe control capabilities in complex environments.

Google DeepMind introduced an inference-time guidance mechanism into Gemini Robotics 2.5, enabling the robot to adjust its actions in real time while executing tasks to avoid surrounding physical hazards.

Results show that this mechanism increased the success rate of safely completing tasks near obstacles by 33%, and by 43% near people.

According to Google DeepMind, Gemini 4 Argon played an important role in the development of Gemini Robotics 2.5.

Robots need large amounts of real physical data to learn new skills, and Gemini 4 Argon can act as a simulation engine, building realistic digital twin environments based on real scene images and generating thousands of different simulated practice runs. Combined with real training demonstrations, these simulation data can help robots adapt more quickly to new scenes and objects.

In addition, Gemini 4 Argon can also infer the reasons for task failures based on videos of robots executing tasks, and test improvement plans in simulated environments and on real robots.

In addition to accelerating robot training, Gemini 4 Argon also performed well in safety evaluations.

Google DeepMind this time open-sourced the safety evaluation benchmark ASIMOV-Jailbreak, which contains hundreds of test prompts covering dozens of scenarios, mainly used to test models' ability to resist disguised dangerous instructions and jailbreak attacks.

In the evaluation, Gemini 4 Argon's jailbreak attack resistance rate reached 99.5%, higher than the previous Gemini Robotics ER 2's 82%.

Conclusion: Large models accelerate training, but real-world operation remains key

This upgrade of Gemini Robotics 2.5 not only reflects the progress of VLA models in robot manipulation capabilities, but also demonstrates more possibilities for general large models in robot training and R&D.

Using large models to generate simulation data and analyze the causes of task failures can improve the efficiency of robots learning new skills and reduce dependence on real-machine data. However, continuously improving task success rates, reliability, and safety remains key for embodied intelligence to achieve large-scale deployment.

Source

Zhidongxi