Google DeepMind says its new robot AI model is 30% to 100% more successful than its predecessor and tied a bow in tests, but the company's own factory pilot still leaves a 4% failure gap.
Google DeepMind on Tuesday introduced Gemini Robotics 2.5, which it calls its most capable vision-language-action model to date. The company says it evaluated the system across more than 200 distinct tasks, including serving food with tongs, pipetting a liquid and tying a bow on a gift box, and reports a 30% to 100% higher success rate than previous baselines depending on the task, with tasks completed 20% to 30% faster. The model works with several robot types, including Apptronik's Apollo 2 humanoid and the Franka Duo platform. It is not generally available and is rolling out to robotics partners, which Ars Technica lists as Agile Robots, Apptronik and Boston Dynamics.
To test the model outside the laboratory, DeepMind partnered with Apptronik and Mercedes-Benz to run Apollo 2 humanoid robots on active automotive production lines in Kecskemét, Hungary, and Sindelfingen, Germany. In what DeepMind describes as a proof-of-concept, the robots performed kitting, sorting metal brackets from bulk bins into containers in the order needed for each vehicle, and reached a 96% success rate. Ars Technica notes that this is probably still short of what a manufacturer needs for repetitive work, and quotes DeepMind director Kanishka Rao as saying that closing the remaining 4% gap will be a steep road. The model also still runs on cloud-based servers, so gains in accuracy and speed come with some latency.
DeepMind also reported safety results and a new test. It says inference-time reasoning guidance lets the robot steer around hazards, improving safe task completion by 33% around obstacles and 43% around nearby people. The company is releasing ASIMOV-Jailbreak, an open-source benchmark of prompts that disguise unsafe requests, in the hope of giving the industry a shared way to test physical AI. DeepMind adds that its Gemini 4 Argon model helped speed development by generating simulated practice runs and digital twins from real camera scenes. All performance figures come from DeepMind's own evaluations, and no independent testing has been reported.
