On May 31, the KU Leuven LICT center organized a half-day workshop on autonomous systems. A good occasion to learn about the research of LICT members, get inspired by the vision of keynote speakers and go in discussion with audience and speakers.
Trustful autonomous systems
Autonomous systems “observe-decide-act” in unstructured environments without human interaction. They hold enormous societal and economical promises, but those promises will not happen until one can provide the assurance that these systems are safe and perform as intended. Safety assurance of autonomous systems is very challenging because of the complex, open context and the use of new technologies for which safety assurance techniques do not exist yet. For example, an unlimited range of potential situations, the self-adaptive behavior due to artificial intelligence, functional insufficiencies, liability and legal issues, they all play their part. Systems fail due to internal failures in decisional layers and because of uncertainties in perception, decision or action functions. Also unexpected adverse situations caused by for example environmental conditions, human actions or other systems, play a role.
Safety assurance of autonomous systems is very challenging because of the use of new technologies for which safety assurance techniques do not exist yet.
In his keynote, Prof. Guiochet (LAAS-CNRS) stated that trust in autonomous systems should be built on three layers: (1) social acceptance criteria, (2) structured arguments and (3) assurance (building) blocks, based on dependability techniques, as used in the safety critical domain. Being active on the latter, he further elaborated on his own research with examples of fault tolerance, fault removal & fault forecasting solutions for autonomous systems.
Prof. Pissoort (M-Group) introduced the European Training Network on Safer Autonomous Systems (SAS). This project aims to train fifteen PhD students in designing safer autonomous systems. The focus of the project is on the design of inherently safe autonomous systems, providing evidence for autonomous systems and providing assurance strategies. The research will be nurtured by case-studies in automotive, robotics, health, oil/gas, IoT, aircraft, agriculture and nautical…
Prof Holvoet (imec-DistriNet) presented the Tumato Software Development approach to design safe autonomous systems that are guaranteed sound, complete and fast. The approach allows to specify the system’s behavior and safety requirements using the concept of states, actions and goals, validate the description and turn it into an execution policy. It was already applied successfully in the SafeDroneWare project.
With her presentation on “Location Privacy and Adaptive Systems”, Dr. Guerses (COSIC) addressed (future) location-based autonomous systems. These systems base their decisions on our patterns of activities, but might show manipulative power and risk being influenced from the outside. Finally, Dr. Emanuilov (CiTiP) addressed the topic from a legal point of view focusing on the liability question. He introduced concepts like “data as a product”, “liability for potential defects”, “state-of-the-art defence” and “ecosystem liability”.
Autonomous systems’ technologies
After the break focus was on the enabling technologies. Prof. Kraft (MICAS) started with his research on sensors and actuators, addressing high performance and smart inertial sensors, new sensing principles like coupled resonators, improved optical and ultrasound sensors. Dr. Vinogradov (Telemic) focused on wireless communications in aerial autonomous systems. He claimed the suitability of LTE technology, optimized by a smarter use of antenna’s, smart positioning and aerial base stations. Dr. Oramas introduced the PSI-research on turning input images into a steering angle by means of deep learning technology. Findings include the superior performance of regression networks, the limited impact of image color and significant gains for stacked frames. Additionally, he discussed the use of artificial data. Being the last LICT speaker Prof. Van hamme (PSI) focused on turning continuous speech directly into meaning, without intermediate textual descriptions, for increasing interaction robustness. Experiments have proven the feasibility of the approach irrespective of language, dialect, voice impairment, idiosyncratic vocabulary and its adaptability to the end-user.
And what about the humans?
Finally, before enjoying a fresh pint of beer, prof. A. Meschtscherjakov (University of Salzburg) addressed autonomous driving from the driver’s point of view, highlighting the new type of “questions” that come into play when humans’ responsibility turns from actual driving into take-over in case of failures of the autonomous system. Crucial will be to keep humans alert, skilled and informed in order to allow for a timely, smooth and safe take-over.