Gemini Robotics 2.5 release notes

Overview of Gemini Robotics 2.5

Gemini Robotics 2.5 is Google DeepMind’s most capable vision-language-action (VLA) model to date, designed to increase the breadth of skills robots can acquire and the speed of that acquisition. By leveraging Gemini 4 Argon as a simulation engine to accelerate research, the model improves performance across more than 200 distinct tasks, enabling robots to handle delicate objects and navigate unstructured real-world environments with greater precision.

Enhanced Dexterity and Task Breadth

Gemini Robotics 2.5 demonstrates significant improvements in task precision, robustness, and speed compared to previous baselines. The model has been evaluated on a diverse set of tasks, including serving food with tongs, tying bows, and pipetting liquids. These evaluations show a 30% to 100% increase in success rates and a 20% to 30% reduction in task completion time. The model maintains compatibility with various hardware platforms, including the Apptronik Apollo 2 humanoid and the Franka Duo platform.

Real-World Industrial Reliability

In a proof-of-concept trial, Gemini Robotics 2.5 was deployed on Apptronik’s Apollo 2 humanoid robots at Mercedes-Benz production lines in Hungary and Germany. The robots successfully performed automotive kitting—the process of sorting parts from bulk bins into organized trays—achieving a 96% success rate. By utilizing the Gemini Robotics embodied reasoning model to track workflows and the VLA model to control physical actions, the system demonstrated the ability to adapt to unpredictable environments and correct for errors in real-time.

Safety and Steerability

To facilitate safer interaction in everyday spaces, Gemini Robotics 2.5 introduces inference-time reasoning guidance. This feature allows robots to steer around physical hazards, resulting in a 33% improvement in safe task completion around obstacles and a 43% improvement when operating near humans. Additionally, the model shows improved resilience to adversarial inputs; it achieved a 99.5% jailbreak resistance rate on the new open-source ASIMOV-Jailbreak benchmark, a significant increase from the 82% baseline established by previous versions.

Accelerated Development via Gemini 4 Argon

Development of Gemini Robotics 2.5 was accelerated by Gemini 4 Argon, which functions as a simulation engine for robotics research. Argon enables researchers to generate thousands of simulated practice runs from real-world camera footage, creating digital twins that allow robots to generalize to new scenes more efficiently. Furthermore, Argon assists in diagnosing task failures by analyzing video footage and testing potential improvements in both simulation and physical hardware.