“The world doesn’t stand still and wait for the robot to make a decision – it has to happen here and now. Robots that navigate in traffic or in other public spaces must be able to make decisions in real time. The uncertainty from the surroundings is often mathematically neglected in the decision processes that control current autonomous systems”, says Olov Andersson, newly promoted doctor within WASP, the Wallenberg AI Autonomous Systems and Software Program.
Measuring uncertainty
It was this deficiency in dealing with uncertainty that he wanted to correct when he started working on his doctoral project at the Division for Artificial Intelligence and Integrated Computer Systems in the Department of Computer and Information Science. The reason that uncertainty is not included in the calculations is that it is not only complicated but also highly demanding, requiring extensive computing power. In addition, a small robot such as a drone is limited by the size and weight of processors and batteries it can carry.Olov Andersson has investigated three different methods, which he has taken from artificial intelligence, machine learning and automatic control.
The first method, which has so far proved to be the best (or at least the most practical), is to use what is known as “probabilistic machine learning” combined with traditional optimisation-based control. The method gives a clear measure of the uncertainty, and the robot or drone learns to navigate safely and make sensible decisions as it flies.
Demonstration at Gränsö
Olov Andersson demonstrated this within the framework of WARA PS, the WASP arena for public safety, at Gränsö Manor in the autumn of 2019. The drone rapidly learned thatThe second method that he investigated made use of neural networks and the type of artificial intelligence known as “deep learning”. This method is based on teaching a neural network which decisions are to be made, when and in which situation. Olov Andersson shows that the method can have advantages when it comes making fast decisions, but it requires huge amounts of training data. Further, it is more difficult to satisfy safety requirements.
“At the moment it requires a simulator to obtain sufficient training data, but I hope that neural networks can become more efficient in the future”, he says.
In the third method, he focussed on allowing the robot to learn more complex internal models of reality in real time, while at the same time having to act. One of the articles included in the thesis describes how this variant can be used to search for survivors after a disaster. It is also possible to include any other information available, such as observations, geographical data, etc.
Cross-disciplinary
“Autonomous systems is a field that needs both cross-disciplinary expertise and the opportunity to carry out experiments. The thesis contains articles that deal with artificial intelligence, robotics, and automatic control. It has been a huge advantage that I was part of WASP and the WASP graduate school. It gave me access to both deep and broad expertise in all these fields for discussions”, he says.
Olov Andersson is to continue his research career with Roland Siegwart, renowned professor in autonomous systems at ETH in Zürich.
“I have been awarded a Wallenberg postdoc through WASP, and I’m really looking forward to getting deeper into the field, such that my research can benefit society in the future.”
In addition to funding from WASP, parts of the research have been carried out within the strategic investment within IT and information technology, ELLIIT, and with funding from the Swedish Foundation for Strategic Research, SSF.
The thesis: Learning to make safe real-time decisions under uncertainty for autonomous robots, Olov Andersson, Department of Computer and Information Science, Linköping University 2020 Supervisor: Professor Patrick Doherty.
Olov Andersson successfully defended his thesis on 29 April 2020.
Video from Gränsö 2019
Translated by George Farrants.
