Google DeepMind used Gemini 4 Argon as a simulation engine to accelerate the development of Gemini Robotics 2.5, which posted a 96% success rate in a parts-kitting proof-of-concept on active Mercedes-Benz production lines4,5. Apptronik's Apollo 2 humanoid robots autonomously sorted automotive parts from bulk bins into sequenced trays at facilities in Kecskemét, Hungary, and Sindelfingen, Germany3.
Gemini Robotics 2.5, a vision-language-action model that translates visual input and natural-language instructions into physical robot movements, was detailed in a Google DeepMind blog post on November 101. Across more than 200 evaluated tasks — ranging from serving food with tongs to transferring liquids with pipettes — the model achieved 30% to 100% higher success rates than previous baselines, depending on task type, and completed activities 20% to 30% faster, according to Google.
Factory workflow and hardware
In the Mercedes-Benz demonstration, two models divided the workflow: Gemini Robotics ER 2 tracked overall progress while Gemini Robotics 2.5 controlled the humanoid's arms and hands. The tested Apollo 2 configuration used a five-finger SharpaWave hand with 22 degrees of freedom. Apollo 2 picked metal brackets from supply bins and placed them into sequencing containers, correcting any gaps. Google said the work could help automate repetitive and physically demanding tasks while allowing factory technicians to focus on more complex assembly work.
Apptronik CEO Jeff Cardenas said the models improved significantly within a few weeks of post-training. Gemini Robotics 2.5 is currently rolling out to Google's robotics partners — including Apptronik and Mercedes-Benz — and works across different robotic platforms such as Apollo 2 and the Franka Duo system.
Safety benchmarks and simulation pipeline
Google DeepMind also reported gains in safe operation: a 33% increase in safe task completion around obstacles and 43% near people in internal evaluations. The company published ASIMOV-Jailbreak, an open-source benchmark built from 100 scenarios and 1,500 adversarial prompts to test whether robotic systems can resist disguised unsafe instructions. Gemini 4 Argon blocked 99.5% of attacks under ASIMOV-Jailbreak, up from 82% for Gemini Robotics ER 2. The benchmark's multi-turn adversarial evaluation contains 845 samples spanning 15 strategy groups and permits up to 10 steps. The repository provides evaluation code and data under an MIT license.
Gemini 4 Argon also played a role in accelerating robot training by generating thousands of simulated practice runs from real-world camera footage and creating digital twins of physical environments. In a separate experiment documented in an October 13 developer post, Argon reconstructed a calibrated physics simulation from a single color photograph of a real tabletop; Google reported the resulting real-robot success rate rose from 56% to 75% and completion time fell from 35 seconds to 24 seconds. The model can also derive hypotheses about error causes from video recordings and test improvements in simulation before deploying them on the actual robot.
Separately, Google widened access to Gemini 4 Carbon — its top-tier model for coding, cybersecurity, and scientific research — to developers, businesses, researchers, and the public worldwide. Google said Gemini 4 Carbon topped benchmarks including CursorBench 4.0, Perplexity AI's WANDR, and Zapier's AutomationBench.
ANALYSIS The Mercedes-Benz trial marks a concrete step from lab demonstrations to automotive production environments for Google DeepMind's robotics stack, though the proof-of-concept scope — parts kitting rather than full assembly — keeps the deployment narrow. The simultaneous release of ASIMOV-Jailbreak and the Gemini 4 Argon safety numbers positions Google to set the terms of the safety conversation for embodied AI before competitors publish comparable benchmarks.