DeepMind Launches Gemini Robotics 2.5 and ASIMOV-Jailbreak Safety Benchmark

Google DeepMind has announced Gemini Robotics 2.5, an update to its robotics-focused AI models, alongside an open safety benchmark called ASIMOV-jailbreak. The company said the new model is up to twice as successful at some tasks, indicating improved performance in robotics-oriented AI workloads. The ASIMOV-jailbreak benchmark is designed to assess safety and robustness, providing a public standard for evaluating potential failure modes or harmful behaviors in robotic systems. The release combines a capability improvement with a safety evaluation tool, underscoring DeepMind’s emphasis on testing advanced AI systems intended for embodied robotics applications.