Robots and AI: 10 still unanswered questions

In this blog post, I look back on my doctorate in computer vision, AI, and robotics, but above all I look to the future: What challenges will we continue to face in this dynamic field in the coming years? Ranging from the appearance of future robots to the internal structure of intelligent systems and their integration into everyday life and society: ten key questions about robots and artificial intelligence that are still unanswered.

by Lukas Twardon

Given the growing enthusiasm for ChatGPT and similar AI models, you might get the impression that we are just a few updates away from hyper-intelligent robots – just put the AI in a fancy case and the smart helper will relieve us of all the stress. But reality looks a little bit different.

Just over four years ago, I completed my PhD thesis which addressed the problem of teaching robots how to handle deformable objects such as items of clothing. The idea of being able to use such a robot in daily life presented me with numerous fundamental but also very practical questions. Most of these questions have still not been conclusively answered.

1. Smart robots or smart objects?

I love the film series Back to the Future. That's why I even started my dissertation with a scene from part two: Marty McFly is seen putting on self-tying shoes (with so-called power laces) and a jacket with sleeves that are way too big at first but, at the push of a button, automatically adjust to the arms of the wearer. In another scene, when the main character falls into a pool and gets wet, the jacket switches to a self-drying mode. So in this vision of the future, no robots are needed at all because the objects themselves are intelligent. In my PhD project, I nevertheless pursued the robot approach. In this way, everyday objects can remain as they are and everyone can decide for themselves whether or not they want to wear digitized garments in the future. But what stops us from developing smart objects and smart robots?

2. Many specialists or one robot for everything?

Robots are the future. So will we all have a whole battery of robots at home in a few years – one mowing the lawn, one doing the cleaning, one playing with the children, and another bringing us breakfast in bed? Or will there be one robot that can do everything? And what would such a general-purpose robot look like? An obvious approach would be to make this robot human-like (humanoid) and – like the robot hand shown above – equip it with five fingers, because many tasks and objects in our environment are made for the human hand. One disadvantage of humanoid robots, however, is that they generally do not have abilities that go far beyond those of humans: They cannot fly, be used as surgical robots inside the body, or self-destruct.

3. Does intelligence need a body?

I mentioned it at the beginning: according to many scientists, ChatGPT on wheels is not an intelligent robot. But why not? The reason for this is the so-called embodiment theory. It states that cognition often requires physical interaction with the environment. When we humans move around, we constantly take in new information. In addition, we change our environment through motor activity, which in turn changes and possibly improves perception. As part of my PhD project, I conducted experiments in which a robot used this strategy, known as interactive perception (see figure above): By grasping the garment and lifting it a little, the fabric deforms in a way that makes it easier for the robot to recognize the opening.

4. How autonomous should our robots be?

In the Robocup, robots compete against each other in soccer as well as in the disciplines of rescue and household robots. The most important rule: once the whistle has blown, the developers are no longer allowed to intervene and the robots act and react completely autonomously to sometimes unpredictable events. This distinguishes them from remote-controlled robots, but also from industrial robots that repeatedly execute pre-programmed movements with high precision. Examples of autonomous robots that are able to navigate their environment by themselves are robotic vacuum cleaners and – in a broader sense – self-driving cars. In a narrower sense, a robot is only truly autonomous if it supplies itself with energy or even sets its own goals (see question 9). But do we humans want such robots?

5. Must robots always learn everything from scratch?

Symbolic AI uses rules and symbols to solve problems, while subsymbolic AI relies on machine learning and lower-level information processing. For decades, both paradigms existed in parallel in the field of artificial intelligence and were considered practically on par. Today, artificial neural networks that are trained with extensive data sets dominate. This method has achieved great success, but also has its downsides, including a lack of transparency, high computational costs, and potential dangers when training and testing with robots in the real world. An alternative strategy is to combine different AI approaches. For example, when handling items of clothing, topological prior knowledge (a pair of pants has three openings) can be modeled explicitly, while the robot learns complex motion patterns implicitly.

6. Do robots need a sixth or seventh sense?

Every intelligent robot relies on sensors to perceive its environment. Three of the traditional five human senses – hearing, sight, and touch – are simulated by robots using microphones, cameras, and tactile sensors, such as those integrated in the fingertips of the robot shown above. In addition, there is proprioception, i.e. the robot's self-perception. Most robots cannot yet smell or taste, but experiments are already being carried out with biological sensors from real animals. The debate about whether and which other sensory modalities robots need is a hot topic. In practice, many robots use ultrasonic or infrared sensors to perceive distances. However, Elon Musk recently switched to a purely camera-based system in Tesla's self-driving cars, reportedly even overruling the concerns and recommendations of his own engineers.

7. What are robots allowed to do?

The question of what robots may or may not do is as old as robotics itself. As early as 1942, Isaac Asimov set out the following laws of robotics in his short story "Runaround":

1. A robot may not injure a human being or, through inaction, allow a human being to come to harm.
2. A robot must obey the orders given to it by human beings except where such orders would conflict with the first law.
3. A robot must protect its own existence as long as such protection does not conflict with the first or second law.

Military robots in particular do not comply with these rules which are purely ethical in nature. The increased use of robots in everyday life will also lead to more and more legal disputes in which the question of guilt will be at the center of attention after humans have been harmed. Possible parties to be held responsible in such cases could be the user of a robot, the manufacturing company, individual developers, and even the provider of a training dataset. It remains to be seen how the legal situation will change as technological progress continues.

8. Will we trust future robots?

Uncanny Valley is the term used to describe the effect that robots that appear very human-like, yet unnatural, often cause discomfort. In 2016, I had the opportunity to attend a robotics conference in Tokyo to which Masahiro Mori, the then 89-year-old discoverer of the Uncanny Valley, was invited. He was joined by dozens of scientists from various disciplines to discuss the factors that influence whether or not robots are perceived as trustworthy by us humans: from design to certain types of behavior and language to gender attributions. So whether we will enjoy living and working with robots in the future depends on much more than just how efficiently and flawlessly the robots will function.

9. Do robots need their own goals?

Machine learning is essentially the stochastic optimization of an objective function. Therefore, choosing a good objective function (or cost / fitness / reward function) has a significant influence on the behavior of modern robots and AI models. It is usually specified explicitly by the developers or, more rarely, derived from the behavior of human experts. Examples of such objectives include continuing a dialog in a way a human conversation partner would according to training data (ChatGPT), and to play soccer in such a way that the probability of scoring a goal is maximized. Some goals (e.g. scoring in soccer) are so difficult to achieve that sub-goals must be defined (reward shaping). However, caution is advised: If a robotic soccer player is "rewarded" when it approaches the goal with the ball, but not "punished" when it moves away from the goal, it may only learn to endlessly dribble in circles. The Swedish philosopher Nick Bostrom speculated in a thought experiment (here as a browser game) that a general AI whose only goal was to produce as many paper clips as possible, would even risk the extinction of humanity just to achieve this goal. So we should be careful what we wish for from robots, because we might get it. This is especially true when it comes to the idea of creating robots that set their own goals.

10. Will robots take our jobs?

Yes, robots are the immigrants of the 21st century. No, seriously: If the robots of the future won't take over a significant part of our daily work, I'll be disappointed. After all, the Czech word robota means "work" or "forced labor". There are enough beautiful and meaningful things to do in this world. Why would it be so bad if robots took our paid labor away? This is because of how business is done on the free market, or rather: because of the underlying objective function (see question 9), namely the individual maximization of private property, i.e. money. An alternative objective function would be, for example, to maximize the overall prosperity of society while at the same time minimizing the work to be done by humans. It is an open political question as to whether humanity is striving for a competition of ideas with such an objective function at all, or whether it will continue to cling to the hope that our monetary economy represents a suitable reward shaping for the goal of prosperity for all.