Using Path Planning Algorithms and Digital Twin Simulators to Collect Synthetic Training Dataset for Drone Autonomous Navigation

Abstract

There is a major challenge in collecting real-world data for training the AI Agents of self-driving Cars, Drones, and Automated Guided Vehicles (AGVs). The process is slow and expensive, since the data must be reprocessed and correctly labeled before use. Furthermore, it is difficult to collect data for corner cases, especially dangerous scenarios that lead to accidents. Another challenge is the distribution of data that we use to train the model to guarantee optimal results without fitting problems or training with meaningless or redundant data. In this paper, we present a novel methodology to use path planning algorithms to generate and label the dataset needed for training AI agents. Our methodology is demonstrated with A* path planning algorithm and Microsoft Airsim Drone Simulator, which eases the obtainment of the required data for creating the obstacles grid and provides the tools needed for simulating the drone's movement.