And how are things looking in the area of multidimensional problems?
Roland Siegwart: Multidimensional learning of complex interrelationships requires millions of training examples and greater computing power to the order of several dimensions. Deep learning as it is understood today is not capable of this. In its current form it still needs a defined objective. It’s very difficult to provide this type of definition for complex workflows.
Presentday deep learning algorithms are not yet much more than programs that enable large data streams to be optimized and analyzed. For example, deep learning makes it possible to identify cancerous tumors (output) based on images (input). Computers are better at this than people as they can access and process large volumes of data much more quickly. But the abilities of artificial intelligence (AI) are still very limited as things stand. It’s therefore a very bold claim to extrapolate AI systems that solve structured and narrowly defined problems to robotic systems expected to tackle the highly complex, multimodal problems we encounter in our everyday lives.
If we fail to make great progress in agriculture and the logistics of distribution, large parts of the world’s population will continue to suffer undernutrition and malnutrition. How can robots contribute to resolving this problem?
Roland Siegwart: There’s great potential for robots to be deployed in agriculture. Robots can continuously monitor fields and intervene immediately if, for example, more water or fertilizer is required or pests need to be removed. In the near future, this will help to make agriculture far more sustainable as resources like water or fertilizers will be used to best effect and pesticides will be administered in precise doses. We expect a fraction of the pesticide volumes used today to achieve the same effect, and it should be possible to do much of the work involved in combating pests “mechanically.” At present, around 30% of food is lost before it even leaves the field, while another 30% is lost during distribution and storage.
It has been proved that robots can learn from human beings. But can humans learn from robots?
Roland Siegwart: There are not many lessons people can learn from robots in their everyday lives just now. But people can aim to develop an optimal working relationship with them, as robot and human skills can complement each other. Robots don’t get tired, they can carry out highly precise movements, and they can carry heavy loads. Human beings are unbeatable when it comes to analyzing complex systems, interacting with other people, and generating new ideas.
Filippo Rima (laughs): I’ve come to the conclusion that we can certainly learn something from robots. Discipline, hard work, precision, and the ability to work under pressure are all virtues that we as people could do with a little more of.