Cognitive Functions of an Intelligent Agent in the Context of Problem Solving On the Example of an Autonomous Forklift Truck

The development and deployment of autonomous vehicles in outdoor environments, such as construction sites, require intelligent planning and control systems to tackle emergent obstacles. Achieving a high degree of vehicle autonomy while ensuring appropriate reactive behaviour in dynamically changing environments requires a multitude of cognitive functions.
These AI systems are often developed for specific domains, which means that problem-solving in unfamiliar environments cannot be guaranteed. Issues that appear trivial to humans can pose insurmountable challenges to AI-controlled
vehicles. To ensure the safe and responsible use of these technologies, safety in autonomous operation must be guaranteed. This raises the question of how autonomous vehicles can increasingly solve problems independently while
simultaneously ensuring safety.
Using the representative example of an automated forklift truck, we implemented a Genetic Learning Algorithm for Behaviour Trees, which we previously tested on our benchmark scenario Pac-Man. The use of Behaviour Trees as a planning and control system offers the advantage of making the decision structure of the autonomous forklift truck readable and comprehensible to humans, thereby easing the assessment of safety aspects. Through the Genetic Learning Algorithm, the forklift truck can learn new Behaviour Trees to solve related but unfamiliar problems. This technical approach delivers positive results in our task scenario as long as we pre-specify the problem. However, the forklift truck needed significant inputs to find solutions to new problems, thereby limiting its autonomy.
In order to improve our understanding of the limitations of the forklift truck, we conducted an interdisciplinary analysis of the Cognitive Functions involved in Problem-Solving. This led us to conclude that the forklift truck lacked the Cognitive
Functions of abstraction and comprehension. These Cognitive Functions were taken care of by the designer during specific implementations, which significantly reduced the degree of autonomy of the commercial vehicle in any appreciably
different task environment.
Based on this analysis, we suggest corresponding technical solutions to automate the two identified Cognitive Functions, thereby potentially increasing the autonomy of the forklift truck. By embedding these proposed solutions in the structure of the behaviour trees, safety guarantees can be ensured. In summary, this work contributes to a deeper understanding of AI in the given example by making the Cognitive Functions related to problem-solving explicit, while simultaneously providing an analytical tool for the further development of autonomous systems.