Gemini Robotics 2.5 hits 96% on Mercedes-Benz kitting with Apptronik's Apollo 2

The pilot ran at Mercedes-Benz plants in Kecskemét, Hungary, and Sindelfingen, Germany[1]. The job was automotive kitting: pulling metal brackets out of bulk bins and placing them into sequencing containers in the exact order each vehicle needs. DeepMind split the work across two models. Gemini Robotics ER tracked the workflow, while Gemini Robotics 2.5 drove the hands and arms and corrected for gaps[1]. The Apollo 2 used Sharpa's five-finger SharpaWave hand with 22 degrees of freedom[3].

"Pairing Apollo 2 with Gemini Robotics, 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 Apptronik co-founder and CEO Jeff Cardenas[1].

The 96% figure has limits. DeepMind does not say how many picks it covers, how long the pilot ran, or how the robots' cycle time compared with a human worker's[2].

Across more than 200 tasks, from serving food with tongs to tying a bow on a gift box and pipetting liquid, DeepMind reports a 30% to 100% higher success rate than previous baselines, depending on task type, and 20% to 30% faster completion[1]. It works on different bodies, including the Franka Duo platform with two-fingered grippers[1].

DeepMind credits Gemini 4 Argon with speeding up development. Argon generated thousands of simulated practice runs from real camera scenes, built digital twins, and helped researchers work out why tasks failed from video footage[1].

Inference-time reasoning guidance lifted safe task completion around obstacles by 33% and around nearby people by 43%, according to DeepMind[1]. It is also releasing ASIMOV-Jailbreak, an open-source benchmark of disguised unsafe requests, on which Gemini 4 Argon scored 99.5% jailbreak resistance versus 82% for Gemini Robotics ER 2[1].

Gemini Robotics 2.5 is rolling out to DeepMind's robotics partners only[1]. Google has not announced API access or pricing[2].