For a robot to be able to put on the washing machine or iron clothes, as Elon Musk wants, it first has to understand how the humans around it behave. What does your body language, your looks and your gestures imply. This human-machine interaction is the field of specialization of Iolanda Leite, an associate professor at the Royal Institute of Technology (KTH) in Sweden who has led research on how robots can learn those details without which they will not be able to share our same spaces.
In this interview, Leite speaks to elDiario.es hours before his speech at the European Conference on Artificial Intelligence, whose 50th edition was held this year in Santiago de Compostela. During her talk, the professor, a specialist in how robots can influence group dynamics, especially in educational and inclusive contexts, presented research in which she demonstrates that systems that “learn informed by human knowledge” achieve better results. But to do this, they need training data that is not as easy to obtain as that of artificial intelligences like ChatGPT.
What is a social robot?
These are all robots that will need to interact with or around people. It’s delivery robots that can deliver last-mile packages, or some robot that’s in the hospital trying to put towels here and there, or these more interactive robots that can provide personalized tutoring for kids, things like that. They are the robots that will eventually be around us in our homes and workplaces.
Generative artificial intelligence has exploded in the last two years, with millions of people using these new tools that understand our commands and what we need very well. But for now they are still behind the screens. What do they need to reach physical space?
It’s a field where we’re definitely seeing a lot of progress as well, but they’re still in very restricted environments. They must have some kind of perception, interaction skills, and the ability to interact in a way that is intuitive to people. We need the robot to be able to understand some aspects of the social environment that surrounds it, such as acting in a way that is appropriate for the people who are interacting with the robot or in its environment to foresee what the robot is going to do. If you look at the language for example, it’s very clear that it’s turn-based, right? You give an indication and you receive something in return. This whole aspect of social interaction is something we are still a little far away from. When is it appropriate to interrupt a person? When should I, as a robot, stop listening to you and start processing what I need to respond appropriately?
If we look at language models [herramientas como ChatGPT]they have a lot of data on the Internet with which they can train and learn in that aspect, how interaction occurs through a screen. But we don’t have that amount of data on human-machine interaction in the real world. We don’t have those simulations. We have videos, but it is difficult to code them to train a robot because they are very intuitive for us humans, but they are still difficult for machines to reason with. We’re talking about the simplest things, like understanding whether or not it’s okay to cross someone’s path, when to interrupt, how to ask for help if they get stuck. These are details like that that we work on.
How can you get that type of data?
The challenge is that the data sets we use are not that large and the way we collect them is often by bringing people into the lab or taking the robots somewhere where we have a controlled environment. That means it is not a scalable method. But, of course, now there is also industry that is looking to collect this data. They’re testing things like robots that are teleoperated by a human, and that are trying to do things that the human would do. So they are trying to collect data on a larger scale, but a lot of data is needed to achieve the same success that we are seeing in language and other modalities of AI. The problem with the robot is that there are many different modalities: it is not just the language, it is the vision, it is the speech, it is the whole social aspect, the context. So you need more complex data and more complex algorithms to put a product like that around people.
Humans driving robots?
AI is very autonomous, of course, but it’s only very autonomous and very impressive because there are a lot of humans behind it encoding data and fine-tuning it. There is a lot of invisible human work behind these systems. One of the things we are working on, although it may seem counterintuitive because we focus on the human aspects of AI, is reducing the human work behind it because it is a lot of burden. Many of these coding jobs are done by people who don’t even know the purpose of what they are doing. That is why there is also the ethical question that automation does not mean generating other, even more precarious, human jobs in less developed countries.
In which sectors do you think we will begin to see this type of social robots?
When you think about the opportunities, I would say it will depend quite a bit on the country and what sectors we don’t have enough people who want to work there. For example, in the medical sector it is usually difficult to recruit people, so it can be an opportunity. But it’s not about replacing people in those jobs, it’s about doing certain things so that humans can focus on providing appropriate care. Things like material deliveries could easily be replaced and I think that would be the first thing. But it is difficult to predict and will depend a lot on the context.
Are you worried about people’s reaction to these robots? Faced with such visible automation?
Well, change won’t happen suddenly. We will not wake up one day with robots around us, but it will be gradual. We already have vacuum cleaners or robotic lawnmowers. If the next step is a robot that can run the dishwasher for me or iron my clothes, I think the reaction will be quite positive. If it’s about replacing these tasks that people don’t want to do. Then there is always the question that eventually some jobs will change. That is a more complicated aspect to address, but I don’t think it will be something sudden, but rather similar to what has happened with new language models, for example. First people find it very funny, then they get used to it and its use gradually becomes normal.
Is this something experts are currently working on? Or is this an area that will be addressed when we get to that point?
No, it is definitely being studied now. In my main area of research, human-robot interaction, there are already many people who are researching these ethical aspects or simply how to center the development focus on the user. It’s not about, ‘Oh, let’s build a robot,’ but about the robot being able to solve some real needs of people in a specific environment. So there is definitely a lot of care with all this and not to fall into technological determinism. Of course, that doesn’t mean there aren’t some Elon Musks in the world doing things their own way.
Musk is also trying to develop an android that he says he will release in the coming years. What do you think of those promises?
I think they are making good progress. Of course, I think he is very optimistic about the challenges, but he has to be because he is a businessman and he has to sell the idea. Maybe by doing that, you’re increasing the risk that then, if what you promise isn’t delivered, there will be another robotics winter or something. On the other hand, it is also good that you set the bar at that level, since people will try to surpass it. It’s not just Tesla that is working on it. There are many other companies now researching humanoid robots and trying to implement them in the real world.
Is the future of robotics that machines look like humans? Androids like the one Elon Musk develops?
There is a good reason for these robots to be human-shaped. And it’s because if we want them to be in our space, we don’t want to have to adapt the spaces to the robot. The spaces are designed for us, with doors of a certain size, tables of a certain height. So if robots have a human form, it is easier for them to do the same things that humans do.
His research addresses how robots can maintain long-term relationships with humans. What characteristics must it have for a person to want to interact with it repeatedly?
For robots to foster long-term relationships with people, they must have features that make them useful and attractive. My research focuses on how to make robots socially competent and effective in educational or therapeutic applications, where their role goes beyond performing specific tasks. For example, in learning, a robot must be able to keep children engaged, motivating them to continue practicing skills such as chess or mathematics. We don’t yet know precisely what specific characteristics foster this connection, but they are likely similar to those we see in human relationships. A good model is a teacher who knows what the student has done well or poorly in the past and adjusts his or her support accordingly. Thus, the robot could adapt its behavior to offer a feedback personalized, encouraging continuous learning and long-term interest.
In therapeutic applications, such as physical rehabilitation, a robot can also play a valuable role as a personal trainer. Many patients must perform repetitive exercises and often become demotivated when they have to do them alone at home. A robot acting as an exercise buddy could offer constant support, reminding you of exercises, providing encouragement and helping to maintain engagement over time.
As you say, these interactions often depend on very small details in which the robot must understand a social aspect of personal communication. Could you explain a specific example of how this aspect is worked on?
There is a great example and it is very simple. We showed that a robot, just by looking at different people at different times, was able to balance people’s participation in a conversation. We had this game where a native Swedish speaker and a person learning the language were trying to guess a word for the robot. Typically, there is an imbalance in that interaction because the native speaker dominates the interaction. So we created this imbalanced scenario to really understand that the robot, ignoring the native speaker, could ensure that both people were included in the interaction. We showed that with an autonomous robot that did this based on some reinforcement learning algorithm, we could balance participation. This was just a game, but it can be used to manage group interactions.
We also have an application similar to that in schools where we have these groups of children who again, native Swedish speakers and children who have recently arrived in Sweden. With a simple steal we showed that we created a more balanced participation in the game. We are trying to analyze whether robots can support social interactions in a positive way, so we are trying to use the robot as a tool to improve social interactions between people.
#Iolanda #Leite #robot #expert #real #challenge #understand #social #context