Google DeepMind says Gemini Robotics 2.5 can complete dexterous physical tasks more successfully and quickly than its previous model, but factory testing shows how far general-purpose robots remain from human reliability.
The system combines an embodied reasoning model, which uses instructions and live video to plan, with a vision-language-action model that generates robot movements. Across more than 200 activities, DeepMind reported success-rate gains of 30% to 100% and speed gains of 20% to 30%. Demonstrations included serving food with tongs, using a pipette and tying a bow.
In European Mercedes-Benz factories, Apptronik Apollo 2 robots using the model kitted automotive parts with a 96% success rate. DeepMind acknowledged that closing the remaining four-point gap is difficult because repetitive industrial work demands near-perfect consistency.
The model still runs on cloud servers, adding latency, and published video was accelerated to twice normal speed. DeepMind also released an open safety benchmark called ASIMOV-jailbreak to test whether robotics models can be manipulated into unsafe actions. Laboratory results and a limited factory trial are useful evidence, but they do not yet demonstrate dependable unsupervised operation across varied workplaces.