Google DeepMind unveils Gemini Robotics 2.5 to speed up robot learning

At a glance

Google DeepMind has introduced Gemini Robotics 2.5, a new AI model that helps robots learn more tasks faster in real-world settings. It is a VLA (vision-language-action) model, meaning it understands sight and language and turns them into physical actions. It is currently rolling out to robotics partners first.

Key points

1. 1

   The model was tested on more than 200 tasks, including serving food with tongs, tying a bow on a gift box, and pipetting liquid.

2. 2

   Compared with earlier models, its success rate is 30% to 100% higher depending on the task, and it finishes tasks 20% to 30% faster.

3. 3

   In a Mercedes-Benz production line test, a humanoid robot sorted and packed car parts with a 96% success rate.

4. 4

   The team improved physical safety and created a new benchmark that tests robots against adversarial jailbreak attempts.

Why it matters

Robots may increasingly take on simple, repetitive work in factories and everyday spaces, so this progress matters for how people work alongside them.