Google DeepMind says a humanoid sorted car parts at Mercedes-Benz with a 96% success rate
Apptronik's Apollo 2 ran Gemini Robotics 2.5 in a pilot at two Mercedes-Benz plants. DeepMind gives no count of parts or attempts.
On 10 November 2026 Google DeepMind released Gemini Robotics 2.5, a model that turns camera images and instructions into robot movement. It says the model drove Apptronik’s Apollo 2 humanoid in a proof-of-concept on Mercedes-Benz production lines in Kecskemét, Hungary and Sindelfingen, Germany. The robot did “kitting”: it took metal brackets from bulk bins and placed them in containers in the order each vehicle needs. DeepMind describes the work as autonomous and reports a 96% success rate.
DeepMind also reports tests on more than 200 tasks, such as serving food with tongs, tying a bow on a gift box and pipetting a liquid. Against “previous baselines” it says the model had 30% to 100% higher success rates, depending on the task type, and finished tasks 20% to 30% faster. Its safety guidance, which steers the robot around hazards while it runs, raised safe task completion by 33% around obstacles and 43% around nearby people.
DeepMind says Gemini 4 Argon resisted 99.5% of the jailbreak prompts in ASIMOV-Jailbreak, a new open benchmark of requests that disguise unsafe robot actions. Its earlier model, Gemini Robotics ER 2, resisted 82%. The model is rolling out to DeepMind’s robotics partners.
What it means
A humanoid doing one sorting job at two car plants is a step beyond lab demos, and 96% is close to what Figure says it needs for car parts: its own BMW target was over 99% placement success per shift. Without the number of attempts and the length of the pilot, a reader cannot tell whether 96% would hold over a full working week.
Caveats
- Every number comes from Google DeepMind’s post, with no independent check. Mercedes-Benz and Apptronik are partners, and Apptronik’s chief executive is quoted in the post.
- The post does not say how many parts or attempts the 96% covers, how long the pilot ran, how many robots took part, or how often a person stepped in. It gives no baseline success rates for the “30% to 100%” comparison. A “success rate” here may not mean the same as Figure’s per-shift placement figure.
- A proof-of-concept is a trial, not production work. Figure’s 90,000 parts at BMW over 11 months is a longer record, also self-reported.
- The post links no technical report for the robot results, and the jailbreak score is for the language model, not the robot in the factory.