Hate Ironing? There’s a Machine Learning-Powered Robot for That

by Jamie Beckett

It’s not as unpleasant as scrubbing the toilet. It’s not as smelly as cleaning the cat litter box. But the tedious task of ironing ranks among the most reviled household chores.

You’re still on your own when it comes to cleaning the porcelain throne or scooping up after kitty. But one day you could hand the laundry over to a machine learning-powered ironing robot, developed by researchers at Columbia University, that can iron shirts, skirts and more.

Robots can easily manipulate rigid objects like coffee cups or computer parts. But they’re often confounded by soft, flexible objects, like clothes, that change shape as they’re handled.

“Robotic ironing is a very challenging task,” said Yinxiao Li, lead author on a paper set to be published at the IEEE International Conference on Robotics.

To solve the problem, the Columbia researchers spent three years building an autonomous ironing pipeline that uses GPU-accelerated machine learning to teach robots each step along the way — how to pick up a garment, recognize it, lay it out to iron and then fold it.

This was a new approach to solving the ironing problem. The researchers first simulated different types of clothing and laundry tasks. When they had enough simulation data, they used it to train the robot, said Li.

GPUs Provide Real-Time Results

When the ironing robot goes to work on real clothing — a shirt, for example — it picks up the item at a random point and rotates it 360 degrees to expose it to a Microsoft Xbox Kinect sensor. The sensor captures the shape and color of what is still an unidentified garment and uses that information to create a digital reconstruction.

Researchers used a GPU to speed up the simulation, training and reconstruction processes. “GPUs give us real-time results,” Li said. “With the help of the GPU, we can reconstruct the garment at the same time we’re rotating it.”

From there, the robot identifies what it’s holding as a shirt by matching the reconstruction to the simulation. With that information, the robot knows how to position the shirt for ironing — using more sensors to detect remaining wrinkles — and, later, for folding.

Beyond the Laundry Room

Li’s latest paper is the last of four describing each step in the ironing pipeline. He conducted this research while in pursuit of his Ph.D. at Columbia, and has moved on to another daily task that many people find a chore: He’s working on self-driving cars at Google.

Unfortunately, a robot butler that can iron is still too slow and costly to be practical for your wrinkled shirts. But its ability to handle floppy, unpredictable objects has lots of applications outside the laundry room, Li said.

Once robots can master handling flexible objects, they could do industrial jobs that involve ropes or cable harnesses, or they could be useful in food production, according to Peter Allen, a professor of computer science at Columbia and co-author on the paper.