DeepMind Unveils Clever Gemini Robotics AI Model
Google DeepMind has introduced Gemini Robotics 2.5, an AI model that lets humanoid robots handle delicate tasks like sorting parts and tying bows.
Can you imagine asking a robot helper to wrap a birthday present, tie a neat ribbon bow, and hand it to you safely? On 10 November 2026, scientists at Google DeepMind unveiled the Gemini Robotics AI model (version 2.5) alongside the Conference on Robot Learning in Austin, Texas. This clever programme gives walking robots the power to understand spoken words, look at objects through camera eyes, and move their fingers with surprising care.
For a long time, factory robots have been big, stiff machines bolted to the floor that could only do one simple action. But this new vision-language-action model changes that completely. It allows tall, two-legged humanoid robots to step into human workspaces and handle fragile items without dropping them or making a mess.
What happened?
DeepMind tested the new software inside busy car manufacturing lines run by Mercedes-Benz in Sindelfingen, Germany, and Kecskemét, Hungary. During these real-world trials, a humanoid robot named Apollo 2, built by the robotics company Apptronik, went straight to work. Apollo 2 sorted automotive parts by picking metal brackets out of bulk bins and placing them neatly into trays, completing this task with an impressive 96% success rate!
To handle the parts so smoothly, Apollo 2 used a special five-fingered mechanical hand called the SharpaWave, which features 22 degrees of freedom. In laboratory trials, researchers tested the AI across more than 200 distinct tasks. The robot proved it could complete gentle, tricky actions, such as serving food using kitchen tongs, measuring tiny drops of liquid with pipettes, and tying bows on gift boxes.
Compared to earlier systems, this new model finished actions 20% to 30% faster and improved task completion by 30% to 100%. Safety was also a top priority for the engineers. The robot improved safe task completion by 33% around obstacles and by 43% when operating right next to humans. DeepMind also launched an open-source safety benchmark called ASIMOV-Jailbreak on Hugging Face to test that the robot cannot be tricked into unsafe behaviour.
Why does it matter to you?
Have you ever tried picking up a tiny grain of rice or tying your school shoelaces? You might do it without thinking, but teaching a machine to use fingers gently has been one of the toughest puzzles in science! In India and across the world, scientists want to build friendly helper robots that can one day assist with everyday chores, like tidying up toys or serving snacks safely.
Because this system helps robots work carefully around people without bumping into them, it brings that helpful future one step closer. Helpful bots might not just stay inside car factories. In the future, they could become reliable helpers in schools, libraries, and homes, carrying out daily activities with great care.
Words to know
- Humanoid robot: A robot built to look and move like a human, usually with two arms, two legs, and a head.
- Degrees of freedom: The number of different directions and ways an artificial joint or robotic finger can bend and turn.
- Vision-language-action (VLA): An artificial intelligence model that translates camera pictures and spoken language directly into physical motor movements.
- Benchmark: A standardised test used by scientists to measure how well and safely a computer programme performs.
Did you know?
Apollo 2’s SharpaWave robot hand has five fingers and 22 degrees of freedom, which allows it to bend, twist, and grip objects almost like a real human hand!
Think about it
If you had a gentle robot helper at home, what is one tricky chore you would ask it to do for you? Why is it so important for engineers to test robot safety before letting machines work near people?
Quick quiz
What success rate did the Apollo 2 robot achieve when sorting car parts in factory trials?
- 75%
- 96%
- 100%
See the answer
96%
Sources
This story was prepared by the Kids Gazette news desk with the help of AI research and writing tools, using the sources listed above.
