Google DeepMind introduces Gemini Robotics 2.5, its most capable vision-language-action (VLA) model to date. This model accelerates real-world general dexterity for robots across everyday skills and factory floors. Testing across 200 distinct tasks showed significant improvements in task precision, robustness, and speed. The model boosts dexterity, reliability, and safety, enabling robots to step out of the lab into unstructured environments.
In a real-world proof-of-concept with Applitronik’s Apollo 2 humanoid robot on Mercedes-Benz production lines, Gemini Robotics 2.5 achieved a 96% success rate in picking and packing automotive parts. The model was also tested with Franka Duo platform featuring two-fingered parallel grippers. Applitronik’s Apollo 2 features a five-finger, 22 degree-of-freedom Sharpa Wave hand, and the model supports both humanoid and parallel-gripper robot types.
The advancement marks a leap in three key areas: Dexterity, Reliability, and Safety. Compared to previous baselines, Gemini Robotics 2.5 achieves 30% to 100% higher success rates depending on task type, while completing tasks 20% to 30% faster.
Source: Google DeepMind Blog · Summarized by HeadlinesBriefing