Google DeepMind unveiled Gemini Robotics 2.5, a new generation of AI-powered robotics platform showing significant improvements in dexterity and real-world task execution. The model demonstrates up to twice the success rate on various tasks, with robots successfully performing complex maneuvers like tying bows and handling deformable objects. In Mercedes-Benz factories, Gemini-powered Apollo 2 robots achieved a 96 percent success rate in automotive parts kitting, though human workers still achieve near-perfect execution.
The system shows 30-100 percent higher success rates and 20-30 percent faster performance compared to previous versions. DeepMind partnered with Apptronik and Mercedes-Benz for real-world testing, utilizing a simulation engine called Argon to accelerate development through digital twins. The platform addresses the challenge of limited physical training data by combining synthetic and traditional demonstrations.
DeepMind Robotics head Carolina Parada notes a trade-off between specialization and generalization, acknowledging that closing the remaining performance gap with humans will be challenging. The open-source safety benchmark ASIMOV-jailbreak aims to ensure safe human-robot collaboration as the technology advances toward broader deployment.
Source: Ars Technica · Summarized by HeadlinesBriefing