Google Releases Gemini Robotics 2.5, Combining an Action Model, Simulation Training, and Robot Jailbreak Evaluations

Google DeepMind released Gemini Robotics 2.5, a vision-language-action model, on November 10, focusing its improvements on dexterous manipulation, task learning, and reliability in real-world settings. The company says the model was tested on more than 200 tasks, including using tools, tying bows, and pipetting. Depending on the task, success rates improved by 30% to 100% relative to previous baselines, while completion speed increased by 20% to 30%. These are vendor-reported comparisons and should not be taken to mean that every robot deployment will achieve the same gains. Official announcement

The more deployment-relevant example is a proof of concept at a Mercedes-Benz factory: an Apptronik Apollo 2 robot removes metal parts from a bin and places them in a container in assembly order. The company reports a 96% success rate. Gemini Robotics ER tracks the workflow, while Robotics 2.5 controls the arm and hands, assigning multi-step reasoning and physical actions to separate models. A single success rate, however, is not enough to assess production-line benefits; engineering teams would also need the number of trials, frequency of human intervention, and recovery time after failures. Trial and architecture details

For training, DeepMind says it used Gemini 4 Argon to create digital twins from camera scenes and generate simulation exercises, alongside real-world demonstration data. On safety, the company added inference-time guidance and announced ASIMOV-Jailbreak, which tests unsafe requests disguised as ordinary instructions. The reported 99.5% jailbreak resistance applies to Argon; it should not be treated as a physical-operation safety rate for Robotics 2.5. Technical details

The existing ASIMOV v2 research already includes risk recognition, safety reasoning, and intervention decisions in its evaluations, providing context for this update. However, it should not be conflated with the newly announced jailbreak test. Researchers should track the new test's data and scoring details, as well as reproducibility across settings and hardware. For now, the company says the model is rolling out to robotics partners. ASIMOV v2 research page