DeepMind reveals Gemini Robotics 2.5 and open safety benchmark "ASIMOV-jailbreak"
The new model is up to twice as successful at some tasks.
TL;DR
- Gemini Robotics 2.5 demonstrates improved dexterity and speed, with success rates up by 30-100% and speed increases of 20-30% in lab tests.
- Robots powered by Gemini Robotics 2.5 achieved a 96% success rate kitting automotive parts in Mercedes-Benz factories.
- The platform uses a simulation engine, Gemini 4 Argon, to accelerate AI training by generating thousands of practice runs based on real-world data.
- Inference-time reasoning helps Gemini Robotics 2.5 avoid hazards, increasing task completion success by 33% with obstacles and 43% with people present.
- DeepMind has introduced ASIMOV-Jailbreak, an open-source safety benchmark with 1,500 prompts designed to test the safety of robotic AI systems.