New to this? Read it in simple words
  • Google DeepMind says Gemini Robotics 2 can control a robot’s whole body and handle objects. It says the model can also plan tasks that take several minutes.
  • Long tasks give small errors time to grow. An object may move or a tool may be missing.
  • A useful robot must notice problems and recover safely. People must also have a quick and obvious way to stop it.
  • Each robot body has different sensors, reach and strength. These differences still need testing.
Words to know
AI model
A computer program trained on lots of data to write, see, speak or make decisions.
Sensor
A device that measures something, like light, touch or distance.

A useful robot must do more than perform one perfect action. It must notice problems and recover safely when the world changes.

A robot must check its own progress

Google DeepMind says Gemini Robotics 2 can control a robot’s whole body, handle objects, and work with other robots. It can also plan tasks that take several minutes and contain many steps.

A home or factory robot must do more than pick up an object. It must check whether the action solved the right problem. If the room changes, it must continue, recover, or ask for help.

Long tasks give small errors time to grow. An object may move, a tool may be missing, or a person may change the instruction. The robot therefore needs to track its progress while it acts.

Whole-body control is also difficult. A robot may need to reach, balance, look, and hold an object at the same time. Safe movement requires constant small changes.

Sources1

Loop
Useful robots check their progress when the world changes.

Long tasks require a robot to sense, plan, act, check the result, and recover after a problem.

Safety needs more than learned AI

An earlier Gemini Robotics release named three goals: broad skills, interaction, and careful hand movement. It also described several safety layers. Normal control software still manages collision avoidance and force limits.

This mixed design is important. Learned AI can understand a messy room. Fixed safety controls can keep the robot inside physical limits that should never change.

A robot that uses a new cup is not automatically safe in every new building. Each robot body has different sensors, reach, strength, and possible failures. These differences still need testing.

Some AI can run directly on the robot. This reduces delay and helps when the internet is weak. However, the model must then work with limited power and computing capacity.

Sources2

People need clear behavior

People do not experience a robot as a test score. They notice whether it stops near a child, recovers after dropping something, and explains why it cannot continue.

A robot should show what it plans to do before it moves. It should explain uncertainty in clear words. People must also have an obvious and quick way to stop it.

General-purpose robots will not arrive all at once. Progress will come through more tasks with clear limits, human supervision, and safe recovery after mistakes.

Sources12

Sources

Every fact in this story comes from the sources below. Open them to check our work.

  1. 1
    Primary source · July 30, 2026Gemini Robotics 2 brings whole body intelligence to robots Google DeepMind
  2. 2
    Primary source · March 12, 2025Gemini Robotics brings AI into the physical world Google DeepMind
How we checked this story

Performance and skill statements are company claims unless the linked paper shows otherwise.

More on this story

Mood: 6.8 out of 10 · Slightly positive

Google DeepMind says Gemini Robotics 2 can control a robot’s whole body and plan long tasks, helping robots recover from mistakes. How we rate the mood

Basics explainedWhat are world models and JEPA?