Gemini Robotics 2.5 - Gemini Robotics 2.5: Google’s most capable robot model

Gemini Robotics 2.5 is Google DeepMind’s most capable vision-language-action (VLA) model to date. In plain terms, it turns what a robot sees and what it’s told into hand and arm movements, with the aim of teaching robots more skills, faster, in less structured real-world settings.

What’s new, per DeepMind: a 30% to 100% higher success rate than previous baselines, depending on the task, and tasks completed 20% to 30% faster. It was evaluated across more than 200 tasks, including serving food with tongs, tying a bow on a gift box and pipetting liquid. It works with tool use, deformable objects like cloth and ribbon, and delicate items, and it runs across robot types, including Apptronik’s Apollo 2 humanoid and the Franka Duo two-arm platform.

Some results from DeepMind’s charts, comparing Gemini Robotics 2 with 2.5:

• Packing pastries with tongs: 5% to 100%

• Closing a toolbox lid: 0% to 93%

• Untying ribbon bows: 0% to 70%

• Hanging a cardigan: 0% to 62%

• Folding pants: 0% to 40%

The real-world test: Apptronik’s Apollo 2 humanoids did automotive kitting (sorting parts from bulk bins into ordered trays) on active Mercedes-Benz production lines in Hungary and Germany, reaching a 96% success rate. DeepMind calls it “promising progress” toward industrial use, so it’s a pilot, not a finished deployment.

On safety, DeepMind added reasoning guidance at run time to steer robots around hazards, and reports 33% more safe task completion around obstacles and 43% around nearby people. It also released ASIMOV-Jailbreak, an open-source benchmark of prompts that disguise unsafe requests.

Good fit if you work on robotics or embodied AI and want to see where general-purpose robot models are heading, especially for dexterous tasks and real production lines.

P.S. I hunt the latest and greatest launches in tech, SaaS and AI, follow to be notified → @rohanrecommends

Really cool! But will it work inside the microduck im getting for xmas? 🎄🙏 (ref: https://pollen-robotics.com/microduck/)

The runtime safety guidance is the part that matters most to me. With coding agents a bad command is recoverable; with a robot arm it isn't, so steering around hazards at run time feels like the right layer to get right first. The jump from 0% to 93% on closing a toolbox lid is the number that stuck with me — that's the kind of everyday manipulation where generalization either works or it doesn't. Curious how much of that carries over to tasks the model hasn't seen in training at all.