Google DeepMind released Gemini Robotics 2.5 on November 10, 2026, and backed up the launch with a factory test: Apptronik's Apollo 2 humanoid robot sorted automotive brackets from bulk bins into sequencing trays with a 96% success rate, running on active production lines at Mercedes-Benz plants in Kecskemét, Hungary, and Sindelfingen, Germany.
The task, known as kitting, requires picking individual parts and arranging them in trays in the exact order a specific vehicle on the line needs them. Apollo 2 handled it autonomously, correcting for gaps and misplaced parts as it went. DeepMind said a companion model, Gemini Robotics ER, tracked high-level progress on the task while Gemini Robotics 2.5 itself drove the humanoid's arms and hands, and that the pairing reached that performance after only a few weeks of post-training on Apptronik's hardware.
Why a factory pilot beats a lab demo
Humanoid robot makers run plenty of warehouse and kitting demonstrations, but doing it inside a live Mercedes-Benz plant, on real production schedules, is a tougher bar. DeepMind says Gemini Robotics 2.5 posted success rates 30% to 100% higher than its predecessor, Gemini Robotics 2, across more than 200 distinct tasks evaluated, and finished tasks 20% to 30% faster. Apptronik CEO Jeff Cardenas said the results carried over to the factory floor with little drop-off: "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."
Google's play for the robot brain layer
Rather than building its own humanoid hardware, Google DeepMind is positioning Gemini Robotics as the model layer that multiple robot makers plug into, with Apptronik, Agile and other partners supplying the bodies. That mirrors how Nvidia has pitched its own robot foundation models and puts DeepMind in direct competition for the same role. The approach also lets Google spread validation across different robots and factories rather than betting on one hardware platform succeeding on its own.
Alongside the Mercedes-Benz results, DeepMind released ASIMOV-Jailbreak, an open-source safety benchmark built from hundreds of prompts designed to disguise unsafe requests across dozens of scenes. The company's newest model, Gemini 4 Argon, resisted 99.5% of those jailbreak attempts, up from 82% on the prior baseline, and DeepMind reported a 33% boost in safe task completion around obstacles and a 43% improvement in safe completion near people.
What comes next for factory humanoids
Mercedes-Benz already has a commercial agreement with Apptronik, and the Kecskemét and Sindelfingen pilots suggest that relationship is moving from pilot demonstrations toward tasks embedded in actual production flow. A 96% success rate is strong for a live line, but it still means roughly one in 25 parts needs human correction, the kind of gap that keeps humanoid robots paired with supervisors rather than running fully unattended. DeepMind says Gemini Robotics 2.5 is now rolling out to its robotics partners, setting up a wider test of whether the model's gains hold up across different robot bodies and factory floors beyond the automaker that hosted the first public proof.