Google DeepMind releases Gemini Robotics 2.5 AI for robots

The release builds on Google DeepMind's push to tie robot control to the same multimodal foundation models used for language and vision. Gemini Robotics 2.5 is a VLA (Vision-Language-Action) model, meaning it turns camera images and instructions into robot movements. The company says it validated the model across more than 200 tasks, from serving food with tongs to tying gift-box ribbons, with success rates rising 30–100% and task times falling 20–30% against prior comparison baselines.

The model is designed to work across different robot bodies, including Apptronik's Apollo 2 humanoid and the dual-arm Franka Duo. In a real-world trial, Apollo 2 was placed on active Mercedes-Benz production lines in Hungary and Germany, where it picked metal brackets from parts boxes and arranged them on trays in the order needed for each vehicle, reaching a 96% success rate. Google DeepMind says the aim is to hand monotonous, physically demanding parts preparation to robots so workers can focus on complex assembly.

Safety features were also strengthened: the model now guides motions during inference and avoids nearby hazards in real time. Google DeepMind's evaluation found that the share of tasks completed safely rose 33% in environments with obstacles and 43% when people were nearby. Development also drew on Gemini 4 Argon to generate simulations and digital twins from real camera footage, which the company says helps robots adapt to new environments and objects more efficiently.