Robots are intended to tackle the monotonous or undesirable jobs that we prefer to avoid. Yet, automating tasks like cleaning a bathroom poses significant challenges, especially when considering how to program a robot arm to reach every part of a washbasin. The complexity increases with uniquely shaped basins and the need to apply the right amount of force at specific points.

Rather than exhaustively writing rules or formulas to cover every scenario, researchers at TU Wien have adopted an alternative method: a human demonstrates the cleaning process to a robot multiple times. With a specially designed sponge, these demonstrations enable the robot to learn through observation, equipping it to handle various shapes and surfaces with ease.

This groundbreaking research was showcased at IROS 2024 in Abu Dhabi, one of the leading robotics conferences worldwide. Cleaning is merely a starting point; many other surface treatment activities in the industry, such as sanding, polishing, painting, and adhesive application, rely on similar concepts.

“Capturing the geometric shape of a washbasin with cameras is relatively simple,” says Prof Andreas Kugi from the Automation and Control Institute at TU Wien. “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?”

People learn these things through experience and imitation. “In a workshop, someone might look over the apprentice’s shoulder and say: You need to press a little harder on that narrow edge,” says Christian Hartl-Nesic, head of the Industrial Robotics group in Andreas Kugi’s team. “We wanted to find a way to let the robot learn in a very similar way.”

A unique cleaning tool has been engineered specifically for this purpose: A cleaning sponge equipped with sensitive force sensors and tracking markers was utilized by humans to meticulously clean a sink—but only its front edge.

“We generate a huge amount of data from a few demonstrations, which is then processed so that the robot learns what proper cleaning means,” explains Christian Hartl-Nesic.

This innovative learning process is driven by a cutting-edge data processing strategy devised by the research team at TU Wien. It merges various established techniques from machine learning: Initially, the measurement data undergoes statistical analysis, and these insights train a neural network to grasp predefined movement elements, known as ‘motion primitives.’

The robot arm is then controlled in an optimal way to clean the surface. This groundbreaking learning algorithm allows the robot to clean the entire sink or other items with complex surfaces after training, even though it has only been instructed on how to clean one edge of the sink.

“The robot learns that you have to hold the sponge differently depending on the shape of the surface, that you have to apply a different amount of force on a tightly curved area than on a flat surface,” explains PhD student Christoph Unger from the Industrial Robotics group.

The technology has far-reaching applications across various industries, including joineries where wooden workpieces are sanded, automotive shops for repairing and polishing paint damage, and metalworking facilities focused on welding sheet metal parts. In the future, the robot could be mounted on a mobile base to serve as a helpful assistant in any workshop environment. These robots could potentially exchange their knowledge with one another.

Sensitive information—such as the exact design of a specific workpiece—would remain confidential, yet fundamental principles would be shared to enhance the overall abilities of the robots. This concept is known as ‘federated learning.’ Multiple experiments conducted at TU Wien have demonstrated the versatility of the sink-cleaning robot.

Numerous tests conducted at TU Wien have consistently demonstrated the remarkable flexibility of the sink-cleaning robot, underscoring its potential in modern manufacturing.

Reference:

1. Christoph Unger, Christian Hartl-Nesic, Minh Nhat Vu, and Andreas Kugi. ProSIP: Probabilistic Surface Interaction Primitives for Learning of Robotic Cleaning of Edges.