‘Teaching a robot to fry an egg isn’t as easy as you’d think’
‘AI can’t do half as much as people think,’ says computer scientist and psychologist Roy de Kleijn. He tries to teach robots seemingly easy things, and keeps on discovering how smart human intelligence really is. Three things that computers are no way near doing.
Everyday tasks
‘Pour 200 millilitres of milk into a bowl.’ A task like this, De Kleijn explains, is incredibly complicated for a robot. It needs to know where to find the carton of milk and the bowl, to understand the word ‘pour’ and then has to perform the actions. Whereas a human, even in someone else’s kitchen, will recognise a fridge almost straight away and will always find something that resembles a ‘bowl’: in short, they can do it straight away.
AI is really good at things like working out chess moves and identifying patterns in data. But computers aren’t ‘handy’. ‘We people can do many different things and are very adaptive. Why that is we simply don’t know. We think the key lies in the brain’s architecture: how those 300 billion neurons are connected.’ The question that De Kleijn is working on is: ‘Which architecture would improve a robot’s neural network? Which connections in our brain still make us humans superior to computers in something as apparently simple as frying an egg?’
Adaptation
De Kleijn is trying to train a robot to touch three points in a certain space. If humans do this, they automatically choose the most efficient order, place their hand at the correct angle: in short, they adapt their behaviour to the situation without thinking. ‘It’s really difficult for a robot to replicate that. How you learn this optimisation of behaviour is something we’re still investigating. In a two-dimensional space it’s fairly possible to replicate human adeptness in a robot arm, but add a third dimension and that makes it much more difficult, let alone if you add gravity or friction to the mix. When there’s no theoretical reason why a robot shouldn’t be able to do this.’
Conversation
‘Back in the 1960s they claimed we’d be talking to computers within ten years. Well, anyone who today in 2021 has ever communicated with a company chatbot knows we’re no way near that. In tests where you have to decide if you’re chatting to a computer or a human, even the best computer’s cover is blown within five sentences. And we haven’t even mentioned sarcasm. That’s yet another step further.’
Roy de Kleijn is an assistant professor and expert in the field of (cognitive) robotics and artificial intelligence at the Cognitive Psychology department.