Inside Google DeepMind’s Robotics Lab: From Lunch Packing to Laundry

Hannah Fry visits Google DeepMind’s California robotics lab with robotics director Kanishka Rao. The 2025 tour shows demonstrations of manipulation, language instructions and combining smaller actions into longer tasks.

Rao says improved visual models have made the lab less dependent on controlled backgrounds and lighting than it was several years earlier. He presents this as progress in visual generalization, not proof that all environments are solved.

The team builds robotics systems on vision-language models that can interpret scenes and instructions. Rao says this can help robots respond to unfamiliar visual situations and language.

A vision-language-action model connects visual and language information with physical actions. The team also uses an agent to decompose longer goals into smaller actions, such as deciding what to pack for a trip.

Rao describes a “thinking” component that outputs intermediate reasoning before acting. He says the team has observed better performance in its tasks, but the demonstration does not establish how reliably the system will work everywhere.

The first robot packs a lunch, including manipulating a small bag and food items. Fry and Rao watch it retry and adjust; they emphasize the fine physical precision required.

Asked how the task was learned, Rao says people demonstrated it through teleoperation, controlling the robot so it could learn from examples of the task.

The next demonstration focuses on generalization. A robot receives spoken instructions to move colored blocks, with a Gemini-based layer interpreting the request.

Fry introduces a stress ball the team says the robot has not seen, along with a small container. She asks it to open the container, place the ball inside and close it; the sequence tests new objects and multiple instructions.

Rao explains that an orchestrator can turn a high-level request into shorter instructions for the action model. The aim is to perform a longer task without the user specifying every movement.

In an example, the system looks up local waste-sorting guidance and sorts items. The demonstration illustrates a proposed combination of information retrieval, reasoning and physical action, not a guarantee of correct disposal in every location.

Rao describes two components for this task: a reasoning model orchestrates a vision-language-action model that performs movements. The two systems work together across the sequence.

A humanoid robot sorts laundry into bins while producing intermediate text about its actions. Rao says this demonstration uses one model for thinking and acting, rather than the separate hierarchy used in the previous example.

Researchers then introduce objects they say were absent from training and ask the robot to place them in containers. Fry observes that the robot is slow and sometimes fails, even as it attempts to adapt.

Rao considers these systems building blocks for general-purpose robotics, while expecting at least one more major breakthrough. He points to the amount of physical interaction data needed for training.

More data alone may not settle every issue, he says: reliable task execution and safety remain challenges. The team also hopes to learn from human manipulation videos, though transferring that information to robots is an open problem.

Fry closes by noting both the progress and the limitations visible in the lab: robots remain slow and clumsy in some cases. Her prediction of a coming robotics revolution is a view about the future, not a claim that these systems are ready for all settings.