DeepMind Unveils Gemini Robotics 2.5 and Open Safety Benchmark ASIMOV-jailbreak

Google DeepMind has introduced the 2.5 version of its Gemini Robotics platform, claiming the system is nearing human-level performance on real-world tasks. The update covers Gemini Robotics 2.5, the vision-language-action model that generates robot movements, alongside Gemini Embodied Reasoning, which turns user instructions and live video into task plans.

According to DeepMind, the new model handles dexterity-heavy work far better, succeeding 30 to 100 percent more often and running 20 to 30 percent faster than its predecessor in lab tests spanning more than 200 activities, from serving food with tongs to pipette work. Apptronik's Apollo 2 humanoid, running the software, can now tie a bow with its five-fingered hands-a difficult feat given how hard deformable objects like ribbon are for robots to handle.

Field testing with Apptronik and Mercedes-Benz saw Gemini-powered Apollo 2 robots kit automotive parts in several European factories with a 96 percent success rate. DeepMind director Kanishka Rao acknowledged that closing the remaining 4 percent gap will be a steep road, since humans performing such repetitive work are almost always correct. The team is also preparing an open source safety benchmark, ASIMOV-jailbreak, for a future in which robots work alongside people.

Source: Ars Technica