Making robots socially intelligent

Raymond Cuijpers is associate professor at the department of Industrial Engineering and Innovation Sciences, where he is part of the Center for Humans and Technology and the Social Robotics Lab. He is also involved in the work of EAISI, the new institute of TU/e in the field of artificial intelligence. AI plays an important role in his research. A short introduction in five questions and answers.

What is your key research question?

The main research question of the Social Robotics Lab is how to make robots socially intelligent. Most artificial intelligence is about correctly identifying a pattern and drawing conclusions from that. For example, AI these days is capable of detecting a face and even classifying its emotions. However, the question of how an artificial agent should react when a person it sees is scowling or laughing, is completely unknown. It depends strongly on the context in which this behaviour manifests itself, and a thorough understanding of psychological concepts, like a person’s intentions. This kind of ‘social intelligence’ is still lacking in robots and other artificial agents. To solve this problem we build models of artificial social intelligence and test them with human participants to validate and evaluate their performance. This kind of research is also very much in the spirit of our Center for Humans and Technology: taking the human perspective into account when designing technology.

What is the key challenge in your work?

There are a number of challenges that relate to making an agent socially intelligent.

First, it requires a lot of knowledge about what drives a person, how to understand human behaviour, how people interpret a robot’s  behaviour and so on. This is hard because knowledge from psychology is not detailed enough for direct application in artificial agents. Worse still, it is often not clear whether psychological theories apply to artificial agents in the first place.

Secondly, we have the problem of interpreting another agent/person’s behaviour. Of course, this can be addressed with Deep Learning methods, but that requires huge, annotated data sets, which are typically not available for socially intelligent behaviour.

Thirdly, we have dynamic interaction. Artificial agents should provide social cues to help humans understand their intentions. It is really in the interaction with others that social intelligence makes a difference.

Finally, general intelligence is still missing. Most artificial intelligence is about obtaining knowledge from sensors about the environment in a form that is easy to process. The next stage of processing (what we call: the behavioural layer) then makes smart decisions about which actions to take given a certain knowledge base. The actions taken will affect the environment, which in turn changes perception, thus closing the so-called perception-action loop. In contrast to robots, humans and other biological agents are very good at  this, by actively perceiving the environment in order to acquire information needed to make proper decisions. In my opinion, infinite recursion like we see in the perception-action cycle holds the key to a more general intelligence, which will enable robots to become truly social

What are the practical applications of your research? How does it benefit society?

When robots become socially intelligent, they will be able to function properly in human environments. Many potential robot applications like in health care, education, therapy, civil services would strongly benefit. For example, because of the ageing society, the number of caregivers is steadily decreasing while the number of elderly in need of care is increasing. Robots could assist the caregivers and support the elderly, but only if users accept the robot. This depends largely in how well a robot can interact with a person, and how socially intelligent it is.  It is with reason one of the fastest growing research areas of the last decade.
 

How do your see the development of AI in the future? 

AI will allow devices to detect complex events quickly and reliably, like speech recognition or facial expression recognition. However, interpreting and responding to these events adequately is notoriously hard and led to the downfall of AI as promising research 30 years ago. The presence of big data and much faster computers has addressed the knowledge gap between artificial agents and people, but the more fundamental problems are still the same and unsolved. So no doubt many useful applications will emerge, but without being honest about the limitations people may be disappointed again about the potential of AI. The main risk is that the current hype turns into a hoax, again.

Why should any AI researcher want to work at TU/e?

AI used to be software only, but is now also commonly used in physical devices that can manipulate our world. The impact will be much larger than before and it is therefore appropriate to focus on the intersection of engineering and artificial intelligence research. TU/e does not have a strong tradition in fundamental AI research like some other Dutch universities, but it does have a strong engineering background. In close collaboration with the high-tech companies of the Eindhoven area, we are increasingly developing practical applications of AI, which benefit both industry and society. Thus, it makes sense to invest in the EAISI institute to boost AI know-how and research, in combination with our strong engineering background and with the needs from industry and society.