When Grooten presented his research in five minutes at the Talking Science pitch competition, he held a toilet brush in his hand—a playful nod to household robots that could take over tedious chores, like cleaning the toilet or folding clothes. But behind the joke lies a personal motivation. “It’s my dream to make this a reality,” he says.

“As a kid, I always wondered if unloading the dishwasher could be automated,” Grooten recalls. Twenty years later, part of that future has arrived: many households have a robot vacuum, but robotic help beyond that is still limited. “Even simple tasks are far harder for robots than you might think,” he says.

“Take folding clothes, for example, which is a major area of research,” he explains. He grabs his own sweater and tosses it crumpled onto the table. “See, it’s tricky for a robot to figure out how it all goes together. Every action we humans take for granted, a robot has to learn—and it has to do that for all kinds of different clothing items.”

Reinforcement learning

To turn a robot into a useful helper, you first need to give it a task, Grooten explains. One way to train a robot is by showing it countless examples. You can also build computer simulations that generate many scenarios in a short time, giving the robot more practice.

A common approach is called reinforcement learning. The robot receives positive points when it performs an action correctly and negative points when it makes a mistake. Its goal is to accumulate as many points as possible. Through trial and error, it gradually learns what is expected.

More adaptive

But robots often learn behaviors that only work well in the specific situations they were trained for. If something changes—a garment is folded differently, or the surroundings look different—the robot can get confused. Move it to another room, for instance, and it might be distracted by the wallpaper, a wooden floor, or a TV that’s on.

In the future, household robots may have access to a flood of data: the temperature in each room, what’s in the kitchen cabinets, how often the doorbell rang that day. “But for a specific task—like making a cup of coffee—only a tiny fraction of that information matters,” Grooten says. All the extra input can actually confuse a robot trained with standard reinforcement learning.

The goal, then, is to create robots that can stay focused on their task, even when faced with unexpected distractions. “In my field, we try to make robots more adaptive. The methods we use are called adaptive reinforcement learning,” Grooten explains. This helps robots generalize across different situations and handle new visual backgrounds or environmental sounds. “We want a robot to understand what to do, even in unexpected situations.”

Smart algorithms

To achieve this, Grooten developed algorithms such as ANF (Automatic Noise Filtering) and MaDi (Masking Distractions), which help a robot ignore irrelevant information. ANF filters out environmental noise, while MaDi masks irrelevant pixels by darkening them so the robot isn’t distracted.

With tools like ANF and MaDi, robots can focus on what really matters—even in changing or unfamiliar environments.

Real robots

Grooten first tested these methods in simulations: the robot performed a simple task against a neutral background. Then the researchers swapped in random images to see if the robot could still complete the task. It worked—the robot ignored the background and didn’t get confused.

“The fun part is testing it on a real robot,” Grooten says. “We trained a small robot with a webcam to get as close as possible to a red dot on a screen. Once it mastered that, we replaced the white background with all kinds of images. Even there, MaDi produced good results after just two hours of training.”

“The best part of this research,” he adds, “is that you build an algorithm and then actually see the robot doing what it’s supposed to do. That’s when it all becomes tangible, and I love that.”

PlayStation

During his PhD, Grooten spent six months at Sony AI developing the SPARC algorithm, which allows a robot trained in simulation to perform well in the real world. By systematically varying simulations, the artificial brain learns to respond adaptively to new situations.