Google and DeepMind today introduced Gemini Robotics 2.5, their latest vision‑language‑action (VLA) model designed to help robots learn more quickly and perform a wider range of real‑world tasks. The update aims to improve dexterity, reliability and safety so robots can move beyond laboratory settings into unstructured environments and assist people across more activities.
What’s new in 2.5
Gemini Robotics 2.5 advances the system in three main areas:
- Dexterity: robots can perform a larger set of tasks with finer manipulation, including use of everyday tools and handling of delicate or deformable objects.
- Reliability: in a real‑world proof‑of‑concept on Mercedes‑Benz production lines, the models enabled Apptronik’s Apollo 2 humanoid to pick and pack automotive parts with a 96% success rate.
- Safety: physical safety steerability was improved and a new benchmark was introduced to assess adversarial jailbreaks specific to robotics.
The model works natively with multiple robot platforms, including Apptronik’s Apollo 2 (equipped with a five‑finger, 22 degree‑of‑freedom SharpaWave hand) and the Franka Duo platform with two‑fingered parallel grippers.
Task breadth and measured gains
Because real‑world settings are unstructured, robots must combine precise manipulation with consistent reliability. The team evaluated Gemini Robotics 2.5 on over 200 distinct tasks — examples include serving food with tongs, tying a bow on a gift box, and pipetting liquids. These tasks require advanced tool use, complex handling of deformable items, and delicate manipulation.
Compared with previous baselines, Gemini Robotics 2.5 delivered notable improvements: task success rates rose by 30% to 100% depending on task type, and tasks were completed 20% to 30% faster.
Real‑world reliability test with Apptronik and Mercedes‑Benz
To assess workplace performance, the developers partnered with Apptronik and Mercedes‑Benz and deployed Apollo 2 humanoid robots on active automotive production lines in Kecskemét, Hungary, and Sindelfingen, Germany.
In the proof‑of‑concept the Apollo 2 autonomously performed automotive kitting — the repetitive task of sorting parts from bulk bins into sequencing trays required for each vehicle. Using the embodied reasoning model Gemini Robotics ER to track workflow and the Gemini Robotics 2.5 actions model to control limbs and hands, the robot picked metal automotive brackets from supply bins and placed them into sequencing containers while correcting for gaps. The pilot achieved a 96% success rate, indicating potential for broader industrial use to offload monotonous prep tasks and free technicians for more complex work.
Safety and jailbreak robustness
As robots enter everyday, cluttered spaces, physical safety is critical. Gemini Robotics 2.5 includes inference‑time reasoning guidance that steers robots away from hazards in real time. In evaluations this guidance increased safe task completion by 33% around obstacles and by 43% when working near people.
To assess robustness against malicious or unsafe instructions, the team released ASIMOV‑Jailbreak, an open‑source benchmark with hundreds of prompts that disguise unsafe requests across many scenes. Gemini 4 Argon achieved 99.5% jailbreak resistance on this benchmark, up from 82% for the prior baseline Gemini Robotics ER 2.
How Gemini 4 Argon accelerated development
Gemini 4 Argon’s robotics research capabilities helped overcome data and modeling bottlenecks. Argon can generate thousands of varied simulated practice runs from a real‑world camera scene and build a realistic digital twin. When combined with real demonstration data, these simulations allow faster generalization to new scenes and objects. Argon can also help researchers analyze video to hypothesize why failures occur and test improvements both in simulation and on real robots.
Accelerated by frontier models like Gemini 4 Argon, Gemini Robotics 2.5 represents a meaningful milestone toward physical AGI by improving dexterity, real‑world reliability and rigorous safety for robots intended to work alongside people.
Acknowledgements
The work was developed by the Gemini Robotics team. The announcement acknowledges contributions and leadership from Huizhong Chen, Radu Soricut, Jean‑Baptiste Alayrac, Melvin Johnson, Edward Grefenstette, Claire Cui, Jon Shlens, David Thacker, Zoubin Ghahramani, Clement Farabet, Koray Kavukcuoglu and Demis Hassabis, as well as many teams across Google and Google DeepMind. Partners named for their support include Agile, Apptronik and Mercedes‑Benz.
