Humans need to understand – and trust – robots that learn by themselves

At the Centre for Applied Autonomous Sensor Systems (AASS) at Örebro University, four AI researchers have spent the past few years developing new models for programming and controlling robots. Their aim has been to make it easier for users to understand what is happening “behind the scenes” and to develop robots that learn from their mistakes and adapt to change.

All four researchers are employed by large industrial companies, so their research focuses on the challenges these companies face.

“We’re trying to make it easier for more people to teach robots new tasks. If robots are only controlled by programmers, this limits their applications in everyday life. We’ve shown that a robot can learn a task after a single demonstration,” explains David Cáceres Domínguez, researcher in computer science at Örebro University.

Behaviour trees are one of the tools used to control robots. Such schemes are complex and can be difficult for ordinary users to understand.

“Many models for robot learning today operate in a ‘black box’. The robot learns what to do, but humans cannot understand the decision-making process behind it. This makes it difficult for people to trust what is happening and to correct mistakes. The design of the behaviour tree that controls the robot is crucial,” says David Cáceres Domínguez, who has developed a method for learning behaviour trees based on humans demonstrating, enabling robots to solve problems autonomously.

The challenge is to combine advanced control with the ability to learn and adapt to an unpredictable world. By making robot behaviour easier to oversee and program, while also allowing them to learn from their mistakes, the threshold for everyday use is lowered, for example, for employees in industrial environments.

“It’s possible to combine clear, structured instructions with robots that learn from experience. Traditionally, there’s been a trade-off between systems that are flexible and can learn from mistakes, and systems that are carefully programmed and easy to oversee. In my research, I show that these two aspects can be effectively combined,” says Marco Iannotta, researcher in computer science at Örebro University, who has developed models that allow robots to adapt their actions to, for example, the size of an object or various obstacles, while remaining transparent and predictable.

Another model for control is “planning domains”, which are based on written instructions such as “if this happens, do this.” It can be clearer, but more difficult to overview. When robots are to be used in everyday life, this creates problems and requires adaptation, for example, in the development of self-driving vehicles.

There is one passage I would like to suggest a small adjustment to, in order to better reflect the core focus of my dissertation. In particular, rather than framing the contribution as “combining different systems for controlling self-driving vehicles”, it would be more accurate to emphasise the learning of interpretable representations of system behaviour. A revised English formulation could be:

“By structuring data in a way that enables humans to understand how self-driving vehicles behave, engineers can more easily inspect, verify and correct the systems, while ordinary users understand what the vehicle is doing and why. This means safer traffic, better use of resources and reduced costs for the introduction of self-driving cars,” says Simona Gugliermo, a researcher in computer science at Örebro University, who uses this research in her work at Scania’s self-driving vehicle department.

For example, if a robot can learn what happens when it drops something slippery, it can also work better with humans – and other robots – in an industrial setting or similar environments.

This opens up opportunities for smarter automation, for example, in manufacturing, logistics or healthcare, where robots can take over repetitive or physically demanding tasks while humans can focus on tasks that require knowledge, creativity or consideration.

“It’s possible to get robots to handle unexpected situations independently and to interact with humans even when something unexpected happens. Meaning they can be used more widely in complex environments in industry, construction, mining and much more,” says Paolo Forte, a researcher in computer science at Örebro University who has studied how this can be done faster and more cheaply using general AI models.