Google DeepMind is rolling out Gemini Robotics 2.5 to its robotics partners, saying the model helps robots acquire skills faster and handle a wider variety of physical jobs. In a proof-of-concept trial at Mercedes-Benz production lines, the company says its models helped Apptronik’s Apollo 2 humanoid achieve 96% success when picking and packing automotive components.
The new vision-language-action model is intended to improve how robots use ordinary tools and manipulate objects that are fragile or bendable. Google DeepMind says those improvements cover dexterity as well as reliability, with the aim of extending robot assistance into less structured settings beyond laboratories. Access is rolling out through robotics partners.
Read more: Google Integrates Intrinsic Robotics Platform to Compete With Amazon and Tesla
Google DeepMind evaluated Gemini Robotics 2.5 on more than 200 separate tasks, examining the range of work it could perform, how well it performed and the complexity involved. The company says that evaluation approach could extend to a larger task set.
For factory workers, Google DeepMind presents the Mercedes-Benz trial as a potential way to shift repetitive duties to robots, allowing technicians to concentrate on more complex assembly work. The company also credits Gemini 4 Argon with helping speed its robotics research, following work earlier this year that extended whole-body intelligence to several robot types, including humanoids.
The update also includes changes that Google DeepMind says improve its ability to steer robots toward physically safe behavior. It has introduced a benchmark for testing resistance to adversarial jailbreaks aimed specifically at robotics systems, alongside the model’s dexterity and reliability work.
The release follows Gemini Robotics ER 2, which launched in July as a reasoning layer for planning multistep robot work and directing lower-level action models. That earlier system could track task progress from video and retry unsuccessful steps without restarting an entire workflow. In October, Google Cloud opened Gemini Robotics ER 2 in public preview through Model Garden on the Gemini platform, adding enterprise deployment and security capabilities around the reasoning model.
