Their robot can clean – took just four days to build

TECH

In record time, the researchers taught their robotic arm to mop up spills efficiently. The trick? To train it using the ai language model GPT-4o.

It doesn’t look like much to the world. But on the other hand – who would say no to a helper that automatically cleans water spills and other messes from the floor at home?

Used 3-printed components

A group of robotics researchers at UC Berkley and ETH Zurich in Switzerland have succeeded in constructing a robotic arm that solves that problem – and with relatively inexpensive components that are based on open source code and software available on the open market.

The researchers’ goal was never to build a cleaning robot for consumers. Instead, the group wanted to try programming a 3d-printed robotic arm to be able to identify and clean up spills.

A robot arm of the model SO-100 from The Robot Studio was used for the project. It was built with 3d-printed components – including six servo motors and an adapter card – at a price of 250 dollars, equivalent to roughly 2,600 kroner.

Ai agent controlled the system

To teach the robot arm to perform the movements required to solve the task, the researchers relied on Open AI’s ai model GPT-4o. The language model was supplemented with a framework from the software company Lang Chain AI to create a so-called ai agent.

An ai agent is based on a multimodal model, in this case GPT-40, which can, among other things, process image and sound to perform actions for the user. This allows the system that drives the robotic arm to use sound to receive instructions from the researchers and moving image to identify what is in front of it – a spot on the table.

The AI ​​agent can then coordinate and control the system itself to solve the problem. These include the robot arm’s movement policies, which are a set of rules for how it should move in different situations – such as when a spill is detected in a certain place.

Celebrates open source

By having the ai model take part in around 100 demonstrations, the system was able to be successfully trained in just four days.

– Open source code really democratizes the field of robotics, writes Jannik Grothusen, one of the researchers behind the project, on the platform X.