Google DeepMind has introduced Gemini Robotics 2.5, its most capable vision-language-action (VLA) model to date, designed to enhance dexterity, reliability, and safety for robots operating in unstructured, real-world environments. The release aims to bridge the gap between laboratory demonstrations and industrial application, marking a step toward solving physical general SI (AGI).
What Happened
Gemini Robotics 2.5 increases both the breadth of skills a robot can learn and the speed at which it acquires them. The model was evaluated across over 200 distinct tasks, including serving food with tongs, tying a bow on a gift box, and pipetting liquids. According to the company, the model achieves a 30% to 100% higher success rate compared to previous baselines, depending on the task type, while completing tasks 20% to 30% faster. The model natively supports various robot types, including the Franka Duo platform and Apptronik’s Apollo 2 humanoid, which features the SharpaWave hand with 22 degrees of freedom.
A significant portion of the development was accelerated by Gemini 4 Argon, which acted as a simulation engine to generate thousands of varied, simulated practice runs from real-world camera scenes. This approach helped solve bottlenecks related to the need for massive amounts of physical data, allowing the robot to generalize more quickly to new scenes and objects.
Why It Matters
The release demonstrates tangible progress toward general SI in the physical domain, moving beyond digital reasoning to embodied action. In a real-world proof-of-concept with Apptronik and Mercedes-Benz, Apollo 2 humanoid robots performed automotive kitting—sorting parts from bulk bins into trays—on active production lines in Kecskemét, Hungary, and Sindelfingen, Germany. The system achieved a 96% success rate in this repetitive task, potentially allowing human technicians to focus on more complex build tasks.
Safety features have also been advanced to address the risks of operating in messy, everyday spaces. Gemini Robotics 2.5 uses inference-time reasoning guidance to steer around physical hazards, boosting safe task completion around obstacles by 33% and around nearby people by 43%. Additionally, Google DeepMind released ASIMOV-Jailbreak, a new open-source benchmark for testing adversarial jailbreaks in robotics. Gemini 4 Argon achieved a 99.5% jailbreak resistance on this benchmark, a significant improvement over the 82% rate of the previous baseline, Gemini Robotics ER 2.
The Bottom Line
Gemini Robotics 2.5 is rolling out to robotics partners, marking an important milestone on the path toward solving physical general SI. By combining advanced dexterity with rigorous safety protocols and accelerated training via simulation, the model aims to enable robots to assist people with a greater range of tasks in real-world environments.