Overview
Gemini Robotics 2.5 is a vision-language-action model released by Google DeepMind on November 10, 2026, described by Google as its strongest VLA to date and currently rolling out to robotics partners.
The gains concentrate in three areas: dexterity, tool use and handling of deformable objects. Google evaluated it across more than 200 distinct tasks, including serving food with tongs, tying a bow on a gift box and pipetting liquid. Against the previous baseline, success rates rise 30% to 100% depending on task type, with completion 20% to 30% faster. The start and end points on individual tasks tell the story: packing pastries with tongs goes from 5% to 100%, closing the lid of a black toolbox from 0% to 93%, untying ribbon bows from 0% to 70%, hanging a cardigan from 0% to 62%, folding pants from 0% to 40%. The tasks that started at zero are the substantive change in this generation.
The real external validation happened on a production line. Mercedes-Benz ran a proof of concept at plants in Kecskemet, Hungary and Sindelfingen, Germany, where Apptronik's Apollo 2 humanoid, fitted with a five-fingered 22-degree-of-freedom SharpaWave hand, picked and packed automotive parts: sorting individual parts from bulk bins into organised trays, a repetitive and physically taxing job. It reached a 96% success rate. Two models divide the work, with Gemini Robotics ER coordinating the workflow and Gemini Robotics 2.5 controlling hands and arms.
One detail on the development side is easy to miss: Gemini 4 Argon served as a simulation engine here, teaching robots skills that would otherwise demand large amounts of physical data. Google says Argon substantially accelerated development of this generation, which explains why the line between video models and robotics models keeps thinning.
Safety produced something comparable this round. Google open-sourced ASIMOV-Jailbreak, a benchmark of hundreds of prompts that disguise unsafe requests across dozens of scenes, published on Hugging Face. Gemini 4 Argon resists 99.5% of them, against 82% for the previous baseline Gemini Robotics ER 2. Safe task completion also improves 33% around obstacles and 43% around nearby people.
Availability needs stating plainly. Google's wording is that it is rolling out to robotics partners, with no mention of a public API, pricing or a waitlist. What is publicly reachable is the ASIMOV-Jailbreak dataset and a preview entry for Gemini Robotics ER 2 in AI Studio, and that is the embodied reasoning model rather than this action model. For most developers this generation is readable but not usable.
Key Features
- 200-task evaluation: Covers serving with tongs, tying bows and pipetting among others, lifting success rates 30% to 100% over the previous baseline depending on task type and cutting completion time 20% to 30%.
- Dexterous manipulation: Controls five-fingered humanoid hands and parallel grippers through actions that need fine force control, such as screwing in a light bulb and tying knots, described as whole-body control from feet to fingertips.
- Tool use and deformables: Extends the task range from rigid pick and place to using everyday tools and handling delicate, flexible objects, with named examples including folding clothes, untying ribbons and tying gift box bows.
- Production line proof of concept: On Mercedes-Benz lines in Hungary and Germany, Apptronik's Apollo 2 with a 22-DOF SharpaWave hand reached 96% success picking and packing automotive parts.
- Split with the ER model: Gemini Robotics ER coordinates workflows and breaks tasks into steps while Gemini Robotics 2.5 turns those steps into hand and arm motion, the pair completing line tasks together.
- Gemini 4 Argon as simulation engine: Uses a video model to generate physical demonstrations, teaching robots skills that previously depended on large volumes of real physical data; Google credits it with accelerating development.
- Open ASIMOV-Jailbreak benchmark: Hundreds of jailbreak prompts disguising unsafe requests across dozens of scenes, published on Hugging Face; Gemini 4 Argon resists 99.5% against 82% for the previous ER 2 baseline.
- Physical safety gains: Safe task completion rises 33% around obstacles and 43% around nearby people, giving measurable rather than described safety improvement.
Use Cases
- Automotive part picking and packing
- Tasks that require tools such as light bulbs, pipettes and brushes
- Handling deformables such as folding clothes, untying ribbons and tying bows
- Warehouse sorting and shelving
- Loading and placing in precision assembly
- Robot jailbreak and adversarial testing in safety-sensitive settings
- Two-arm task division and collaboration
Pros
- Success rates up 30% to 100% across 200 tasks, with several previously zero-success tasks entering working range
- Completion 20% to 30% faster, which registers directly on line takt time
- Uses everyday tools and handles deformable objects, widening the task range from pick and place to assembly and tidying
- 96% success on Mercedes-Benz lines, the hardest external validation so far
- Ships with the open ASIMOV-Jailbreak benchmark, taking jailbreak resistance from 82% to 99.5% on comparable terms
- Clean split with the ER model, workflow decomposition and motion control each where they belong
Pricing
Currently rolling out to robotics partners only, with no published API pricing, public application route or waitlist. Publicly reachable are the ASIMOV-Jailbreak dataset on Hugging Face and the Gemini Robotics ER 2 preview in AI Studio, which is the embodied reasoning model rather than this action model.
Summary
Robotics progress this year is measured less by single demos than by how many tasks cross the usability line at once. Across 200 tasks Gemini Robotics 2.5 lifts success rates 30% to 100% over its baseline, pulling a batch of zero-success tasks into working range, and finishes 20% to 30% faster. The 96% on Mercedes-Benz lines is the hardest external evidence so far. Two caveats: that figure comes from a proof of concept on one production line, not a mass-production yield, and this generation is only rolling out to robotics partners with no public API and no published price.
Version History
- Gemini Robotics 2.5 released (2026-11-10): Google DeepMind released Gemini Robotics 2.5. Across 200 tasks success rates rise 30% to 100% over the previous baseline with completion 20% to 30% faster: packing pastries with tongs from 5% to 100%, closing a black toolbox lid from 0% to 93%, untying ribbon bows from 0% to 70%. On Mercedes-Benz lines in Hungary and Germany, Apptronik's Apollo 2 with a 22-DOF SharpaWave hand reached 96% success picking and packing automotive parts, with the ER model coordinating workflows. Gemini 4 Argon served as a simulation engine. Google also open-sourced the ASIMOV-Jailbreak benchmark, where Gemini 4 Argon resists 99.5% of attacks against 82% for ER 2, plus 33% better safe completion around obstacles and 43% around people. Rolling out to robotics partners only.