Gemini Robotics 2.5 accelerates real-world general dexterity

The model was evaluated across more than 200 distinct tasks, including serving food with tongs, tying a bow on a gift box, and pipetting liquid. DeepMind frames these as steps toward human-level dexterity that combine advanced tool use, manipulation of deformable objects, and careful handling of delicate items. Reported gains over previous baselines include success rates 30% to 100% higher depending on task type, and completion speeds 20% to 30% faster. DeepMind credits Gemini 4 Argon with accelerating the underlying robotics research, following earlier work this year that brought whole-body intelligence to robots of many types, including humanoids.

Gemini Robotics 2.5 is designed to run natively across different robot platforms. Among those named are Apptronik's Apollo 2 humanoid, equipped with the five-finger SharpaWave hand offering 22 degrees of freedom, and the Franka Duo platform, which uses two-fingered parallel grippers. That hardware range is meant to cover both fine, human-like manipulation and simpler industrial grasping without separate model variants.

The reliability claims were tested in an active workplace rather than a lab. In a proof-of-concept with Apptronik and Mercedes-Benz, Apollo 2 units worked on production lines in Kecskemét, Hungary and Sindelfingen, Germany, performing automotive kitting — sorting parts from bulk bins into trays in the order each vehicle requires. The models reached a 96% success rate picking and packing automotive parts, which the partners present as a way to offload monotonous, physically demanding prep work so technicians can focus on complex build tasks. DeepMind argues the models let Apollo reason over multiple steps and adapt when conditions change.

Safety is treated as a separate axis of progress. DeepMind reports improved physical safety steerability and says it has introduced a new benchmark designed to test against adversarial jailbreaks specific to robotics, an acknowledgment that robots moving into everyday environments face manipulation risks distinct from conventional software prompts.

What to watch next is whether a proof-of-concept on two automotive lines generalizes: the 96% figure comes from kitting tasks in specific plants, and the source does not say how many robots ran, how long the trials lasted, or how the models perform on other factory jobs. Equally open is how the new robotics jailbreak benchmark gets adopted, and whether the 200-task evaluation suite expands as DeepMind claims the approach could scale to many more tasks.