Folding laundry is a task we take for granted, but for a robot it remains one of the most difficult challenges to automate: it requires recognizing a soft and deformable fabric, understanding its orientation, and applying the correct sequence of movements.
Steven Gong and a team of engineering students addressed this exact problem as a graduation project, building a system that learned to retrieve, iron, fold, and stack laundry items autonomously, within a single day of training.
In the future, could we say goodbye to folding laundry? A robot learned in a few hours — here’s how
The team built a system with four robotic arms: two manually controlled by an operator, whose movements are replicated in real time by the other two motorized arms, equipped with grippers to grasp objects and some 3D-printed components. By moving the main arms, the team collected about 200 demonstrations of the entire process in about four hours of work.
On the AI model front, the first attempts with ACT and Diffusion Policy did not yield satisfactory results. The team then moved to π0.5, a 3.3-billion-parameter model, already trained on a large amount of robotic manipulation data, exposing it to the collected demonstrations.
The path wasn’t linear: moving the workstation altered the lighting and surrounding environment, forcing the team to collect additional data to adapt the system to the new conditions.
Despite these setbacks, by eight o’clock in the morning the next day, about 24 hours from the start of the project, the robot managed to complete the entire process on a T-shirt: retrieval, ironing, side fold, central fold, and transfer onto the stack.
The result, as the team themselves admit, was not “a particularly elegant solution”, but it nonetheless demonstrates how a pre-trained model can be adapted to a specific task with a relatively small amount of data and examples.
For all the latest AI news, you can find our most recent articles in GizChina.it’s AI section.
