Google Unveils Gemini Robotics 2.5: Its Most Powerful VLA Model
Google has launched Gemini Robotics 2.5, its most powerful Vision-Language-Action (VLA) robot model to date, featuring significant improvements in dexterity, reliability, and safety. The model has already achieved a 96% success rate in real-world testing on Mercedes-Benz manufacturing lines using Apptronik's Apollo 2 humanoid robots. This release represents a major step forward in bringing general-purpose robotics out of laboratories and into unpredictable real-world environments like automotive assembly plants. By achieving high task completion rates on complex chores, it demonstrates the commercial viability of deploying autonomous humanoids for repetitive industrial labor. Tested across more than 200 tasks—including delicate deformable object handling and tool usage—Gemini Robotics 2.5 delivers 30% to 100% higher success rates and 20% to 30% faster execution compared to its baseline. It pairs with a reasoning model called Gemini Robotics ER and relies on Gemini 4 Argon for accelerated simulation and synthetic data generation.
## BACKGROUND
Vision-Language-Action (VLA) models are advanced artificial intelligence architectures that combine computer vision, natural language processing, and robotic motor control to allow robots to understand instructions and interact physically with their surroundings. Apptronik's Apollo 2 is an advanced humanoid robot platform designed for industrial automation and heavy logistics tasks.