From washing urinals to tidying up the beach, we can already see a future where our robot servants help keep our world a little cleaner. Now, a robotic arm has mastered the surprisingly complex task of sink washing, showing off its ability to learn.
Cleaning a wash basin might not sound like the most advanced of tasks, but when you think about it, a lot goes into it. You have to intuitively know what angle to employ your sponge at, understand how much force to apply to different parts of the sink based on the grime, and readjust your body constantly as you move along the surface. It's certainly easy for us humans, but if you're a programmer working with a just-starting-out robot, it's a lot to code.
"Capturing the geometric shape of a washbasin with cameras is relatively simple," says Andreas Kugi from the Automation and Control Institute at TU Wien in Austria. "But that's not the crucial step. It is much more difficult to teach the robot: Which type of movement is required for which part of the surface? How fast should the motion be? What's the appropriate angle? What's the right amount of force?"
Understanding that programming all of those data points and combinations was a herculean task, Kugi and his team decided to let their robotic arm learn to do the task by observing someone else doing it.
So they developed a special cleaning sponge equipped with force and position sensors and had a person use it to repeatedly clean just the front edge of a sink that had been sprayed with a dyed gel imitating dirt. They then used the data collected from those exercises to train a neural network that could translate the input into predetermined movement patterns. They fed those patterns to the robot and let them inform its movements as it sets out about the task. As you can see in the following video, the training worked quite well.
Federated Learning
While the experiment was focused on sink cleaning, the researchers say that it demonstrates the fact that the robot arms could carry out a range of tasks on different, varying surfaces including sanding, painting, or welding sheet metal. What's more, they say a fleet of robots could learn the basic moves from each other through what's known as "federated learning" and then they could apply those moves to their individual, specified tasks.
"Let's imagine many workshops use these self-learning robots to sand or paint surfaces," says Kugi. "Then, you could let the robots gain experience individually with local data. Still, all the robots could share the parameters they learned with each other."
Can you say, "singularity?"
A paper describing the team's work is available from TU Wien. It was recently submitted to the IROS 2024 conference and was awarded "Best Application Paper Award," setting it apart from over 3,500 other papers.
Source: TU Wien