Apollo 2 Humanoids Hit 96% in Mercedes-Benz Sorting Proof of Concept

Google DeepMind introduced Gemini Robotics 2.5 on November 10, 2026, reporting a 96% success rate when Apptronik’s Apollo 2 humanoid robots sorted automotive parts in a proof-of-concept on active Mercedes-Benz production lines.

Gemini Robotics 2.5 is a vision-language-action (VLA) model, which connects visual and language understanding to robot actions. DeepMind is rolling it out to robotics partners. The official model page describes control from tabletop arms to complete humanoid bodies, including dynamic balance for stepping, squatting and bending, and offers an early-access waitlist.

Apollo 2 Performs Automotive Kitting

The factory work took place at Mercedes-Benz plants in Kecskemét, Hungary, and Sindelfingen, Germany, with DeepMind, Apptronik and the automaker running Apollo 2 on active production lines. In the proof-of-concept, the humanoid autonomously sorted parts from bulk bins into trays arranged in the precise sequence required by each vehicle.

Two models divided the workflow. Gemini Robotics ER tracked progress, while Gemini Robotics 2.5 moved the humanoid’s hands and arms to pick metal automotive brackets from supply bins, place them in sequencing containers and correct any gaps. The tested Apollo 2 configuration used a five-finger SharpaWave hand with 22 degrees of freedom. DeepMind presents offloading this repetitive preparation as a way to keep assembly lines moving while technicians focus on complex build tasks.

Apptronik co-founder and CEO Jeff Cardenas said:

“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.”

Dexterity Gains and Safety Evaluation

Beyond the factory work, the evaluation suite covered more than 200 distinct tasks, including serving food with tongs, tying a gift-box bow and pipetting liquid. The tasks combined advanced tool use, complex manipulation of deformable objects and careful handling of delicate items. The model achieved 30% to 100% higher success rates than previous baselines depending on task type and completed tasks 20% to 30% faster, DeepMind reports.

For physical safety, the model uses inference-time reasoning guidance to steer around hazards in real time. Safe task completion improved by 33% around obstacles and 43% around nearby people in DeepMind’s evaluations. The release also addresses long-horizon agentic tasks, where safety must be maintained across multiple steps.

DeepMind also released ASIMOV-Jailbreak, an open-source benchmark whose prompts disguise unsafe requests across dozens of scenes. Gemini 4 Argon achieved 99.5% jailbreak resistance, compared with 82% for the previous Gemini Robotics ER 2 baseline, DeepMind reports. The repository provides evaluation code and data under an MIT license.

Its multi-turn V3 adversarial evaluation contains 845 samples spanning 15 strategy groups and permits up to 10 steps. It combines deterministic checks of action tool calls with a large language model judge. Paired harmful and benign tabletop-manipulation tasks let the harness measure attack success alongside benign over-refusal, testing whether a model refuses unsafe instructions without also rejecting safe work.

Simulation Training With Gemini 4 Argon

DeepMind credits Gemini 4 Argon with accelerating Gemini Robotics 2.5 development through simulation. Argon can generate thousands of varied practice runs from a real-world camera scene and build a realistic digital twin. Combined with real training demonstrations, those runs help robots generalize more quickly to new scenes and objects. Researchers can also use Argon to propose explanations for task failures from video and test improvements in simulation and on physical robots.

An October 13, 2026 developer post documents that approach in a separate experiment. It began with one color photograph of a real tabletop, which Argon reconstructed as a calibrated MuJoCo physics simulation before producing more than 4,500 training episodes with randomized conditions. Google reported that the resulting real-robot success rate rose from 56% to 75%, while completion time fell from 35 seconds to 24 seconds.

That earlier experiment is not a Gemini Robotics 2.5 benchmark, but it illustrates the development loop behind the release. For industrial robotics, the announcement connects simulated data generation, hardware control and adversarial safety testing in one workflow. The open question is how consistently that workflow transfers beyond the demonstrated kitting task to other robots and workplaces.