Gemini Robotics 2.5 sorts parts at 96% on Mercedes-Benz lines, with Gemini 4 Argon kept off the robot

Google DeepMind released Gemini Robotics 2.5 on Nov. 10, a vision-language-action model it is rolling out to its robotics partners, and it led with a factory result. In a proof-of-concept on active Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany, Apptronik's Apollo 2 humanoids picked metal brackets out of bulk bins and placed them into sequencing containers in the order each vehicle needs. DeepMind puts the success rate at 96%.

The release shows Google's two-speed robotics stack working on a real line. Robrief's October analysis found Gemini 4 Argon far too slow to steer a robot arm in real time. In 2.5 it stays off the robot. DeepMind says Argon served as a "simulation engine," generating thousands of practice runs and a digital twin from a real camera scene, and helped researchers work out from video why tasks failed. On the Mercedes lines, the embodied-reasoning model Gemini Robotics ER kept track of the workflow while Gemini Robotics 2.5 moved the hands and arms.

Benchmarks measured against DeepMind's own baselines

DeepMind says it evaluated the model on 200 distinct tasks, such as serving food with tongs, tying a bow on a gift box and pipetting a liquid. Against its previous baselines, it reports 30% to 100% higher success rates depending on the type of task, and tasks finished 20% to 30% faster. Apollo 2 runs the model with the five-finger, 22-degree-of-freedom SharpaWave hand, and the Franka Duo platform runs it with two-finger parallel grippers.

Safety got the same treatment. Reasoning guidance applied at inference time raised safe task completion by 33% around obstacles and 43% around nearby people, by DeepMind's evaluation. The lab also released ASIMOV-Jailbreak, an open benchmark of hundreds of prompts that disguise unsafe requests across dozens of scenes. Argon scores 99.5% jailbreak resistance on it, against 82% for Gemini Robotics ER 2.

Kitting forgives little

Kitting is the prep work of sorting individual parts into trays in the exact order the line will use them, and DeepMind says Apollo corrected for gaps as it worked. If the 96% is counted per pick, it means four misses in every 100, each one a bracket out of sequence unless something downstream catches it. DeepMind gave no figure for the same task before post-training, so the release shows where Apollo ended up, not how far it moved.

Apptronik is confident about the move. "We saw the models improve dramatically in just a few weeks of post-training and carry over to Mercedes-Benz production with very little drop-off," said Jeff Cardenas, its co-founder and chief executive.