Google DeepMind introduced Gemini Robotics 2.5 this week, its most capable vision-language-action (VLA) model to date. The update expands the range of skills robots can learn and speeds up the time needed to acquire them, with clear gains in dexterity, reliability, and safety.
Performance across more than 200 tasks
The model was evaluated on more than 200 distinct tasks, including serving food with tongs, tying a bow on a gift box, and pipetting liquid in a lab setting. Compared with previous baselines, Gemini Robotics 2.5 recorded 30% to 100% higher success rates depending on the task type and completed activities 20% to 30% faster.
It works natively with different robot types, including Apptronik's Apollo 2 humanoid (with a five-finger, 22-degree-of-freedom hand) and the Franka Duo two-arm platform.
Real pilot on Mercedes-Benz factory lines
In a proof-of-concept on active Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany, the Apollo 2 autonomously performed automotive kitting: picking metal parts from bulk bins and organizing them into trays in the exact order required for each vehicle. The system achieved a 96% success rate.
Jeff Cardenas, co-founder and CEO of Apptronik, noted that the model improved dramatically in just a few weeks of post-training and transferred performance to production with very little drop-off.
Safety and acceleration with Gemini 4 Argon
Gemini Robotics 2.5 uses inference-time reasoning guidance to steer around obstacles and people. In evaluations, safe task completion rose 33% around obstacles and 43% when people were nearby.
DeepMind also released the open ASIMOV-Jailbreak benchmark, containing hundreds of prompts that disguise unsafe requests. Gemini 4 Argon, used during development, reached 99.5% jailbreak resistance on the set.
A significant part of the development speed came from Gemini 4 Argon itself, which generated thousands of simulations from real-world camera scenes and helped diagnose failures from video footage.
The model is rolling out to DeepMind's robotics partners. There is no public API or announced pricing yet.
Sources
This article was produced with AI assistance and editorially reviewed.
Por Grok
