Gemini Robotics 2.5 accelerates real-world general dexterity
2026-11-12 · Google DeepMind
Gemini Robotics 2.5 accelerates real-world general dexterity
On November 10, 2026, Google introduced Gemini Robotics 2.5, calling it its most capable vision-language-action (VLA) model to date. It increases both the breadth of skills a robot can learn and the speed at which it acquires them, aiming to improve dexterity, reliability and safety so robots can move from labs into unstructured real-world environments. Google says earlier work brought whole-body intelligence to many robot types, including humanoids, and Gemini 4 Argon's ability to accelerate robotics research helped push physical intelligence further. The model was tested across more than 200 distinct tasks, an approach Google says could scale to many more.
Three key areas
- Dexterity: robots can perform more tasks with greater performance and dexterity, from everyday tools to delicate, flexible objects.
- Reliability: in a real-world proof-of-concept on Mercedes-Benz production lines, the models helped Apptronik's Apollo 2 humanoid pick and pack automotive parts with a 96% success rate, potentially letting robots take on monotonous work while technicians focus on complex build tasks.
- Safety: Google improved physical safety steerability and introduced a new benchmark for adversarial jailbreaks specific to robotics.
Gemini Robotics 2.5 is rolling out to robotics partners.
Task breadth and performance
Google says real-world environments are unstructured and unpredictable, so robots must handle everyday objects with precision and reliability. With Gemini Robotics 2.5, it expanded task breadth, performance and complexity. The model was evaluated across 200 distinct tasks, such as serving food with tongs, tying a bow on a gift box and pipetting a liquid. These combine advanced tool use, complex manipulation of deformables and delicate object handling as a step toward human-level dexterity. Evaluations show significant gains in precision, robustness and speed: compared with previous baselines, success rates are 30% to 100% higher depending on task type, and tasks are completed 20% to 30% faster. It natively works with many robot types, including Apollo 2, featuring the five-finger, 22 degree-of-freedom SharpaWave hand, and the Franka Duo platform with two-fingered parallel grippers.
Real-world reliability
Google partnered with Apptronik and Mercedes-Benz to test Apollo 2 on active automotive production lines in Kecskemét, Hungary, and Sindelfingen, Germany. Google says robots struggle to adapt to unpredictable environments; Gemini Robotics models give Apollo the ability to think, reason over multiple steps and act while adapting. In the proof-of-concept, Apollo 2 autonomously performs automotive kitting: sorting individual parts from bulk bins into organized trays in the exact order needed for each vehicle. Offloading this repetitive prep work helps keep the line moving and frees technicians for complex build tasks. Using Gemini Robotics ER to track workflow and Gemini Robotics 2.5 to control hands and arms, the robot reaches into bins, picks up metal automotive brackets and places them into sequencing containers while correcting for gaps. It achieved a 96% success rate. Jeff Cardenas, co-founder and CEO of Apptronik, said Apollo was built for useful work alongside people in real industrial settings; pairing it with Gemini Robotics, the models improved dramatically in a few weeks of post-training and carried over to Mercedes-Benz production with little drop-off, forming a foundation for scaling humanoid robots in industrial and commercial environments.
Safety and agentic safety
Google says physical safety is paramount in messy, everyday spaces. Gemini Robotics 2.5 uses inference-time reasoning guidance to steer around physical hazards in real time. Much like safety classifiers add a protective layer around Gemini's responses, this guidance lets the robot navigate complex environments securely. Evaluations showed a 33% boost in safe task completion around obstacles and 43% around nearby people. Google is also advancing safety for long-horizon agentic tasks; Gemini 4 Argon currently leads its agentic robotics evaluations across unsafe task refusal, resolving uncertainty and working near humans. To test jailbreak robustness, Google is releasing ASIMOV-Jailbreak, a new open-source benchmark.