Google DeepMind releases Gemini Robotics 2.5 for industrial robots

An AI model drove a Mercedes-Benz car-part kitting job on its own. The success rate was 96 percent, Google DeepMind said. The task was carried out by Apptronik's Apollo 2 humanoid robot. Apollo 2 ran a new Google DeepMind model called Gemini Robotics 2.5. Google published the figure in a blog post. The test was a pilot at Mercedes-Benz plants in Hungary and Germany.

Gemini Robotics 2.5 is a vision-language-action model. It takes in camera frames, text and robot state. It returns actions for the robot's joints. Google says it was tested across more than 200 distinct tasks. Those include tool use, soft-object handling and precise picks. The new model reaches 30 to 100 percent higher success than Gemini Robotics 2. The exact range depends on the task. The figure is by Google's own measurement. Task completion is 20 to 30 percent faster.

The safety numbers

Google released a new open-source benchmark alongside the model. It is called ASIMOV-Jailbreak. It tests whether a robot model will carry out a dangerous instruction. The prompt is written in a tricky way. Gemini Robotics 2.5 refuses 99.5 percent of the jailbreak prompts on this benchmark, Google says. The previous baseline, Gemini Robotics ER 2, refused 82 percent. Safe task completion around obstacles improved by 33 percent. Around nearby people it improved by 43 percent. Each number is measured against the previous model.

Who said what

Jeff Cardenas, co-founder and CEO of Apptronik, is quoted in the post. "This is what we built Apollo for: useful work alongside people in real industrial settings," he said. He added that Apollo 2, paired with Gemini Robotics, improved "dramatically in just a few weeks of post-training". The gains carried over to Mercedes-Benz production "with very little drop-off". Apollo 2 has a 22-degree-of-freedom hand called SharpaWave. Agile is named as another partner. Google's Franka Duo platform also runs the model.

What it does not say

Every figure in the post is Google's own. The 96 percent success rate is for one kitting task on one production line, done as a pilot. The post does not say how many robots ran. It does not say how many hours they ran. It does not say how the number changed over the pilot. Mercedes-Benz is quoted as a partner, not as a paying customer. ASIMOV-Jailbreak is open source, so another team can rerun it. The success-rate measurements are not.

A team building a robot product now has a second serious VLA model to pick from. The other options are open models and the earlier Gemini Robotics 2. Picking between them means running each one against your own cell, not against Google's. The ASIMOV-Jailbreak benchmark is the most reusable part of this release. It can be run on any VLA model on a laptop. The previous baseline score is published for comparison.

Source

• Gemini Robotics 2.5 accelerates real-world general dexterity, Google DeepMind