Publikation
Interactive Simulator Framework for XAI Applications in Aquatic Environments
Ahmed H. Elsayed; Tarek Elmihoub; Christoph Manss; Andre Miedtank; Lars Nolle; Frederic Theodor Stahl
In: Max Bramer; Frederic Theodor Stahl (Hrsg.). Artificial Intelligence XLI - SGAI 2024. SGAI International Conference on Artificial Intelligence (AI-2024), 44th SGAI International Conference on Artificial Intelligence, December 17-19, Cambridge, United Kingdom, Pages 144-157, Lecture Notes in Computer Science, Vol. 15446, ISBN 978-3-031-77915-2, Springer Nature Switzerland AG, 11/2024.
Zusammenfassung
Trust in Artificial Intelligence (AI) systems is essential for their lasting success, and methods for understanding and justifying their results are of paramount importance. This paper addresses this need by presenting a simulation framework, where an interacting user is prompted through an interface to describe and explain their actions depending on different situations. This simulation framework can generate a dataset annotated with explanations for training explainable AI models for mission planning. Firstly, This paper presents the development of a simulator built with Unity3D. The simulator recreates an aquatic use case that involves aquatic vessel mission planning for lake maintenance. Therefore, the simulator randomises environmental conditions and simulates various interactions with simulated boats on the lake. Secondly, the paper introduces an annotation interface integrated into the simulator to collect textual actions and their explanations. Here, a skipper of a boat in the lake can describe and explain actions, which are then captured together with the boat camera’s current view. In addition to the captured image, instance and semantic segmentation of the boat’s current view can be recorded as ground truth, along with bounding box annotations of the objects in the simulator. The dataset is then used to pinpoint these explanations in a visual context, i.e. generate grounding visual explanations, through a multi-modal object detector, i.e. MDETR or YOLO-World. The source code and the dataset for explaining the skipper’s actions collected using the simulator is available at github.com/dfki-ni/aqua-sim-xai.