Google robot model hits 96% success on Mercedes-Benz assembly lines

Key Points

1. Google DeepMind launched Gemini Robotics 2.5, its most capable vision-language-action model to date.
2. A humanoid running the model packed car parts at Mercedes-Benz plants with 96% success.
3. The trial moves general-purpose robots closer to paid work on real production lines.

The latest:

A humanoid robot running Google DeepMind’s newest model sorted and packed automotive parts on Mercedes-Benz production lines with a 96% success rate, according to the company. DeepMind said the result came from a real-world proof-of-concept at plants in Kecskemét, Hungary, and Sindelfingen, Germany, using Apptronik’s Apollo 2. The underlying system, Gemini Robotics 2.5, is now rolling out to robotics partners.

Details:

• The model: DeepMind described Gemini Robotics 2.5 as its most capable vision-language-action model so far, saying it widens both the range of skills a robot can learn and the speed of learning. The company framed the goal as moving robots out of the lab into unstructured environments.
• The testing: According to DeepMind, the model was evaluated across more than 200 distinct tasks, measured on breadth, performance and complexity. The company said the tasks ranged from operating everyday tools to handling delicate, flexible objects that typically defeat rigid industrial automation.
• Dexterity gains: DeepMind reported success rates 30% to 100% higher than previous baselines depending on task type, with tasks completed 20% to 30% faster. The company did not publish a breakdown of which task categories sat at the low end of that range.
• The factory job: Apollo 2 performed automotive kitting autonomously, DeepMind said: sorting individual parts out of bulk bins into organized trays in the exact sequence each vehicle on the line requires. Kitting is a sequencing task where a single misplaced part can stall downstream assembly.
• Safety steering: The company said inference-time reasoning guidance raised safe task completion by roughly 33% around obstacles and 43% around nearby people. That second figure addresses the core obstacle to deploying humanoids beside human workers rather than behind safety cages.
• Jailbreak benchmark: DeepMind released ASIMOV-Jailbreak as an open-source benchmark for resistance to adversarial instructions. It said Gemini 4 Argon reached 99.5% jailbreak resistance, against 82% for the earlier Gemini Robotics ER 2 baseline.
• Hardware reach: The model natively supports multiple robot types, DeepMind said, naming Apptronik’s Apollo 2 humanoid and the Franka Duo platform. The rollout is going to robotics partners; the company named no commercial launch date or pricing for broader access.
• Mercedes-Benz status: DeepMind characterized the carmaker deployment as a proof-of-concept rather than a production contract, and did not say how many units were involved, how long the trial ran, or whether Mercedes-Benz plans to scale it.

Background:

Vision-language-action models combine visual perception, natural-language instruction and physical control in one system, letting a robot be told what to do rather than programmed for each motion. The approach is the main bet across the humanoid industry that general-purpose machines can replace task-specific automation.

Between the lines:

The 96% figure is the number that matters commercially, but it was recorded in a proof-of-concept, not sustained production, and DeepMind gave no duration or unit count. Automotive kitting is also a comparatively forgiving test: repetitive, indoors, with parts of known shape. The 43% improvement in safe completion near people indicates where the harder engineering still sits.

What’s next

Watch whether Mercedes-Benz converts the trial into a production deployment, whether DeepMind publishes sustained uptime figures beyond the proof-of-concept, and how rival models score on the newly open-sourced ASIMOV-Jailbreak benchmark.