New method helps autonomous systems deal with uncertainties
Self-driving cars, logistics in warehouses, but also predicting maintenance moments in factories: autonomous systems are used for more and more different purposes in our society. However, autonomous systems do not naturally function safely and effectively. They need to take into account the uncertainties in the surroundings. In collaboration with the universities of Oxford, Birmingham Kentucky and Twente, researchers of Radboud University have developed a method to deal with these uncertainties, based on historical data. They recently received a ‘Distinguished paper award’ from the Association for the Advancement of Artificial Intelligence (AAAI) for their publication.
Autonomous systems have to be able to function safely and efficiently. The circumstances in which autonomous systems operate are often complex and hard to predict. Thom Badings, researcher at Radboud University, explains: “Take for example a drone that needs to safely deliver a package. The drone cannot crash on its way and it needs to have enough fuel to safely arrive at its final destination. External factors, such as wind, cause uncertainty: the wind can influence the position and speed of the drone. Our research is focussed on making safe automatic decisions in such uncertain circumstances.”
Figure 1: Different factors influence the safe arrival of the drone.
Anticipate uncertainty in the surroundings
How do you anticipate uncertainty? In the case of the drone delivery problem, the existing models rely on strong assumptions on the probability distributions of external factors. “But in real life, these assumptions are often not realistic”, according to Nils Jansen, researcher at Radboud University. “Take the wind: you should be able to exactly predict the probability of a wind gust at a certain time and the strength of the gust.”
Contrary to current methods, these researchers do not rely on a precise probability distribution for this prediction of external factors, but rely on earlier measurements or simulations of these factors. In this way, they can guarantee the probability that a drone will safely arrive at its destination, even before it has taken off. Nils Jansen: “Our method is based on historical data of uncertain factors, such as wind. We combine ideas from control theory, artificial intelligence, and formal verification to program the controllers for autonomous systems. Our method may be used to continuously learn from reality, so we can reduce the uncertainty over time.”
Predefined safety threshold
In the case of the drone that needs to deliver a package safely, the researchers work based on a predefined safety threshold. For example: the probability that a drone safely arrives at its destination needs to be higher than 98%.
The wind force can vary, and so the ideal route of the drone can also vary. “With our method, we can take this into account. If there is little wind, the drone can take the short, but narrow, route. If the wind is stronger, it is safer to take the long route that has less obstacles”, Thom Badings explains. “With our method, we can program the controllers that take all these matters into account. This makes them safer than the controllers that were programmed with current methods.”
Figure 2: When there is little wind, the drone takes the short but narrow route (blue). If there is a strong wind, the drone chooses for the long, but safer route (purple).
With this research, the researchers take an important step in designing safe controllers of autonomous systems. Nils Jansen: “In the future, this method can be used in various ways, such as robotics, fintech, or predictive maintenance. To achieve this, we are now working on making our algorithms even faster, and applying our methods in systems with multiple sources of uncertainty.”
- Contact Thom Badings or Nils Jansen
- Publication: Sampling-Based Robust Control of Autonomous Systems with Non-Gaussian Noise, Thom S. Badings, Alessandro Abate, Nils Jansen, David Parker, Hasan A. Poonawala, Marielle Stoelinga.
- Distinguished paper award AAAI for publication Thom Badings, Nils Jansen and Marielle Stoelinga (Radboud University)
- Read: A new model reduces uncertainty in AI (Innovation Origins, January 2021)