~/AI/deepmind-reveals-gemini-robotics-2-5-and-asimov-jailbreak-benchmark

DeepMind Reveals Gemini Robotics 2.5 and ASIMOV-Jailbreak Benchmark

Google DeepMind has announced Gemini Robotics 2.5, delivering up to twice the success rate of previous versions on specific robotic tasks. Alongside the model update, DeepMind introduced an open-source safety benchmark called ASIMOV-jailbreak to evaluate physical AI safety. This release marks a significant advancement in physical AI and robotics performance, bridging digital models with real-world environments like factories. Furthermore, the introduction of the open ASIMOV-jailbreak benchmark addresses critical safety standards for the entire robotics industry. Gemini Robotics 2.5 achieved a 96% success rate in a Mercedes-Benz factory proof-of-concept using Apptronik's Apollo 2 humanoid robot. Additionally, the ASIMOV-jailbreak benchmark features hundreds of prompts disguising unsafe requests across dozens of scenes, demonstrating a 99.5% jailbreak resistance for the new Argon model.

## BACKGROUND

Gemini Robotics models function as Vision-Language-Action (VLA) systems that process visual and textual inputs to directly control physical robots. Safety benchmarks like ASIMOV-jailbreak are designed to test whether AI models can resist malicious prompts that trick them into executing harmful or unsafe physical behaviors.

## REFERENCES

## KEYWORDS

#AI#Robotics#Google DeepMind#Machine Learning#Safety

$ subscribe --daily