Gemini Robotics 2.5 Claims 96% Success Rate in Factory

The new vision-language-action model Gemini Robotics 2.5 is introduced with performance increases measured across 200 tasks and an industrial pilot with the humanoid Apollo 2. The designers also highlight safety improvements, an open-source benchmark for testing evasions, and substantial support from Gemini 4 Argon for simulation.

Claimed Safety Improvements and New Evasion Benchmark

The model's promoters claim to have enhanced maneuverability in physical safety and introduced a dedicated benchmark for adversarial evasions specific to robotics. Gemini Robotics 2.5 applies reasoning guidance to inference to avoid hazards in real-time. According to their assessments, this approach increases the safe completion of tasks by 33% around obstacles and by 43% near people. For longer-duration agentic tasks, they indicate that Gemini 4 Argon tops their internal evaluations in refusing dangerous tasks, managing uncertainties, and working near humans. They are releasing ASIMOV-Jailbreak, an open-source benchmark compiling hundreds of disguised requests posing dangerous demands across dozens of scenes. In this set, Argon shows a 99.5% resistance to evasions, compared to 82% for Gemini Robotics ER 2 on the previous benchmark.

A Pilot on Mercedes-Benz Lines with Apollo 2 Achieves 96%

A real-world test involved Apptronik and Mercedes-Benz on active automotive production lines in Kecskemét and Sindelfingen. The designers remind us that robots struggle to adapt to unpredictable environments and assert that their models provide Apollo with multi-step reasoning and adaptability to changes. In the proof of concept, the humanoid Apollo 2 autonomously performed automotive kitting, sorting parts from bulk bins into ordered trays according to the required sequence. The workflow was monitored by Gemini Robotics ER, while Gemini Robotics 2.5 controlled hands and arms to grasp metal supports and place them into sequencing containers, correcting gaps. The pilot reports a success rate of 96%. According to the promoters, offloading this repetitive work would help maintain the assembly pace and refocus technicians on complex construction tasks.

Claimed Performance Increases Across 200 Tasks

Gemini Robotics 2.5 has been evaluated on over 200 tasks, including serving food with tongs, tying a ribbon around a gift box, or pipetting a liquid. These scenarios combine advanced tool use, manipulation of deformable objects, and delicate handling of fragile items. Internal evaluations report notable increases in accuracy, robustness, and speed. Compared to their previous benchmarks, they indicate success rate gains ranging from 30% to 100% depending on task categories and execution times reduced by 20% to 30%. They add that this approach could be extended to many other tasks.

Multi-Platform Compatibility and Dexterity Advances

On the dexterity front, the designers claim a significant progression, from everyday tool use to the manipulation of soft or fragile objects. They specify that the model works natively with several types of robots, including Apptronik's humanoid Apollo 2 equipped with the SharpaWave five-fingered hand with 22 degrees of freedom, as well as the Franka Duo platform equipped with two-finger parallel grippers.

Large-Scale Simulation with Argon and Announced Roadmap

The promoters attribute a simulation engine role to Gemini 4 Argon, which they claim has significantly accelerated the model's development. According to them, Argon generates thousands of simulations from a camera-captured scene to build a realistic digital twin, then helps researchers analyze failures from videos and test improvements, both in simulation and on real robots. Combined with real-world demonstrations, this approach aims for faster generalization to new scenes and objects. They remind us that they have already deployed complete body intelligence on various robots, including humanoids, earlier this year. Gemini Robotics 2.5 is presented as a step towards physical AGI and is designed to equip robots with the intelligence necessary to work alongside humans on concrete tasks. The deployment of the model with partners is announced to be underway, with stated goals of expanding learned skills, accelerating learning, and improving dexterity, reliability, and safety in unstructured real-world environments. They also describe Gemini Robotics 2.5 as the most advanced VLA model in their range to date.