Leiden University logo.

nl en

Roy de Kleijn: ‘Fetching a glass of water is extremely difficult for a robot’

Training a robot in a real-life environment is more or less impossible. Computer scientist and psychologist Roy de Kleijn switched to training in a simulated, virtual one instead. ‘You do have a reality gap, but the more precisely you train the robot, the better it is in practice.’ His inspiration: the human brain.

Why is it so difficult for a robot to do something as simple as fetch a glass of water in any old kitchen? They can’t do it yet at any rate. Roy de Kleijn: ‘The trouble is that you can’t separate the task from the environment.’ There are different glasses and different taps, and the robot has to recognise and use them all.

Roy de Kleijn with a glass of water
Assistant Professor Roy de Kleijn is an expert in the field of (cognitive) robotics and artificial intelligence at the Department of Cognitive Psychology.

In his PhD research he tried to discover how this kind of practical task could be possible for a robot. Then, in 2017, he took a different tack. ‘I started working with a simulated robot in a simulated environment. The operating models that emerge are definitely applicable in reality.’ He readily admits that you do have a reality gap, but the gap closes the better you train the robot. ‘Take self-driving cars: if you do enough training with large numbers of robots, you end up with a stable model that is reliable in traffic.’ 

He incorporates different AI systems in his virtual environment and lets them train themselves

De Kleijn is researching what a robot ‘brain’ needs to be able to learn how to carry out a complex task, such as moving a particular box in a warehouse. He incorporates different AI systems in his virtual environment and lets them train themselves by interacting with their environment. ‘Which architecture does the robot’s neural network need to achieve this, and how can we make the best network even better?’

Looping in the brain

Imagine a robot that can only open a red door if it first picks up a green block. This only works if it has a kind of looping in its neural network. De Kleijn: ‘We call this recurrence. People can do this kind of thing really easily, which confirms that we too have this looping in the neural network of our brains.’

What does De Kleijn need for the breakthrough that will make robots just as competent as humans? That is the big question. ‘Perhaps we don’t know what the best kind of training is. Perhaps we haven’t chosen the best neural network. We know what a human brain is: 300 billion nerve cells and the connections between them. But what is the secret of human intelligence?’

Humans as a source inspiration, or not?

An important movement within the AI world sees the brain as a source of inspiration. Roy de Kleijn: ‘Not everyone in the AI world believes in this. I don’t know, but today’s deep learning systems, in which a machine is trained with data to differentiate between dogs and cats, for instance, or to detect a tumour, are based on this principle. It’s the work of cognitive psychologists from the 1980s.’

Text: Rianne Lindhout
Photo: Patricia Nauta

This website uses cookies.