Gemini Robotics 2.5: Accelerating Physical Intelligence
Introducing Gemini Robotics 2.5, our most capable vision-language-action model enhancing dexterity, reliability, and safety for real-world environments.
TL;DR
- Gemini Robotics 2.5 is a new vision-language-action model designed to increase the speed and breadth of skills robots can learn.
- Key improvements include enhanced dexterity for handling everyday and delicate objects, increased reliability demonstrated by a 96% success rate in automotive part kitting, and advanced safety features.
- The model was evaluated across over 200 distinct tasks, showing up to a 100% higher success rate and 30% faster task completion compared to previous baselines.
- A real-world proof-of-concept with Apptronik's Apollo 2 humanoid robot at Mercedes-Benz production lines successfully performed automotive kitting, freeing up technicians for complex tasks.
- Safety advancements include real-time steerability around physical hazards and a new benchmark, ASIMOV-Jailbreak, to test resistance against unsafe requests, achieving 99.5% jailbreak resistance.
- Gemini 4 Argon's simulation capabilities accelerated the development of Gemini Robotics 2.5 by providing large amounts of simulated data and efficient modeling algorithms.
- Gemini Robotics 2.5 is rolling out to robotics partners.