Technology news and discussion on TalkFuse.
DeepMind unveils robot AI update and open safety benchmark
Google DeepMind has introduced Gemini Robotics 2.5, a new version of its AI system for controlling robots, alongside an open safety benchmark called ASIMOV-jailbreak. The update aims to help robots handle a wider range of physical tasks, and trials in Mercedes-Benz factories show progress beyond laboratory demonstrations.
Across more than 200 laboratory tasks, DeepMind reports success rates 30 to 100 per cent higher than the previous version, with tasks completed 20 to 30 per cent faster. In factory trials, Apptronik’s Apollo robots achieved a 96 per cent success rate while kitting automotive parts, though DeepMind says matching human-level reliability remains difficult. The system used simulated practice runs based on real-world camera data as well as demonstrations, and still relies on cloud computing.
In short
- Gemini Robotics 2.5 improved performance on dexterity tasks.
- Apollo robots achieved 96 per cent success in factory trials.
- DeepMind introduced an open benchmark for robot safety.
Explained simply
AI systems for robots interpret information such as camera images and help machines decide how to move and handle objects. Unlike software that works only with text or images on a screen, these systems must operate in physical spaces, where objects, people and conditions can vary.
Google DeepMind develops artificial intelligence, while Apptronik makes the Apollo robots used in the factory trials. Mercedes-Benz provided the factory setting for those trials. Results from laboratory tasks and factory work can show how well a system transfers from controlled tests to real workplaces.
Safety benchmarks are sets of tests used to assess how systems respond to potentially harmful instructions or situations. An open benchmark can let other researchers examine and compare results. Reliability matters because robots may work near people, and the reported system still depends on cloud computing.
Read the full article at the source →
Originally published by Ars Technica as “DeepMind reveals Gemini Robotics 2.5 and open safety benchmark “ASIMOV-jailbreak””.