A robot lifting a heavy crate looks impressive. But try asking it to seal a plastic bag. Two hands must hold a floppy object, find the zip seal and apply just enough pressure. With Gemini Robotics 2, Google DeepMind shows robots getting better at these small, awkward actions.
In the demonstration, fingers work together on a knot, a light bulb and a resealable bag. Everyday objects reveal one of robotics’ big challenges: picking something up is different from manipulating it skilfully between your fingers.
Seeing, understanding and moving
Gemini Robotics 2 is the AI that turns camera images and plain-language instructions into movements. DeepMind presented the model on 30 July 2026. The software must decide not only which object is needed, but how arms, wrists and fingers should carry out the task together.
Take unscrewing a light bulb. The fingers must maintain their grip while the wrist turns, then reposition without dropping it. A bag poses a different problem: pull on it and its shape changes, moving the seal. Repeating one fixed movement is not enough.
Five fingers with 22 ways to move
DeepMind tested the same model on platforms including Apptronik Apollo 2 with Sharpa hands. Each Sharpa Wave hand has five fingers and 22 actively controlled degrees of freedom: separate movements for bending, spreading and gripping. More flexibility offers more possibilities, but the software must coordinate them.
Sharpa also builds touch sensing into the fingertips. It can detect pressure, contact and slipping, so the robot need not rely solely on its cameras. That helps when an object disappears behind the fingers or shifts in the grip.
Unscrewing works better than screwing in
This progress does not yet amount to an all-purpose, dexterous household robot. In DeepMind’s own tests, unscrewing a bulb succeeded in 92 per cent of attempts; screwing it in reached 36 per cent. Finding and following the thread remains much harder.
That is why these small actions matter. Homes and workplaces are full of caps, fasteners, packaging and tools designed for human hands. The better robots become at using them, the less their surroundings need to be rebuilt for machines.
