Google DeepMind's Gemini Robotics 2.5 hits 96% success in Mercedes-Benz trial

Putting a robot into a real automotive assembly line is usually a quick recipe for failure. In the real world, lighting changes, parts sit at awkward angles, and people walk by unexpectedly.

Google DeepMind wants to show that its models can finally handle the mess. The lab just rolled out Gemini Robotics 2.5, its most capable vision-language-action (VLA) model to date. To prove it, DeepMind put the system inside Apptronik’s Apollo 2 humanoid robots on live Mercedes-Benz production lines, sorting and packaging car parts with a 96% success rate.

A brain for planning, a body for execution

DeepMind's physical AI platform splits the workload into two layers. Gemini Embodied Reasoning acts as the high-level planner, turning spoken or written instructions and camera feeds into a step-by-step game plan. Gemini Robotics 2.5 is the execution layer, translating that plan into real-world motor controls.

To train the new model, DeepMind used Gemini 4 Argon as an automated simulation engine. Argon turned single camera feeds into realistic digital twins and generated thousands of simulated practice runs, while diagnosing task failures directly from video footage.

In laboratory benchmarks covering more than 200 tasks, Gemini Robotics 2.5 posted a 30% to 100% boost in success rates compared to previous baselines, while moving 20% to 30% faster. The system showed off fine dexterity by tying ribbons into bows using Apollo's five-fingered hands, pipetting liquids, and using food tongs.

Inside Mercedes-Benz plants in Hungary and Germany, Apollo 2 took on automotive kitting. The robot picked metal brackets from bulk bins and sequenced them into assembly trays, adjusting its grip and correcting errors when parts shifted. That is where it achieved the 96% reliability score.

Testing against jailbreaks on the factory floor

When a 150-pound metal machine works around people, physical safety is non-negotiable.

DeepMind added real-time inference safety steering, which helped the model navigate obstacles 33% more safely and boosted safe operation near humans by 43%. Alongside the model, DeepMind open-sourced ASIMOV-Jailbreak, a benchmark with hundreds of disguised, malicious prompts designed to trick robots into dangerous behavior. Tested against this benchmark, the underlying Argon system scored a 99.5% defense rate, up from 82% in the previous generation.

Why it matters

A 96% success rate on an active car line sounds impressive, but it is not quite ready to run an entire shift alone. As DeepMind director Kanishka Rao acknowledged, human workers complete repetitive kitting tasks accurately nearly every single time, making that final 4% gap a steep hill to climb. In automotive manufacturing, a 4% failure rate still means stopped lines and human intervention.

What makes this release clever is DeepMind’s platform strategy. By partnering with hardware makers like Apptronik and Agile Robots rather than building its own humanoids, Google is positioning Gemini as the default operating system for physical machines. Hardware teams can focus on motors, hands, and manufacturing, while DeepMind supplies the reasoning and control layers.

Open-sourcing ASIMOV-Jailbreak also helps Google set the rules of the road. As regulators and industrial buyers scrutinize physical AI safety, having the dominant benchmark gives DeepMind significant leverage over how real-world robotic security is measured.