Dynamic target tracking for quadrotors using liquid neural networks

Abstract

Image Based Visual Servoing (IBVS) and autonomous navigation in robots often rely on phys ical markers for fast detection and estimation of pose and orientation. However, in real-world scenarios, the use of fiducial markers is impractical and infeasible, making marker less navigation a challenging problem. We explore the use of Liquid Neural Networks to achieve generalization and causal understanding for the task of flying towards a target without reliance on markers. By focusing on causality rather than object-specific features, we demonstrate robust performance even in untrained scenarios. We employ a Neural Circuit Policies (NCP), a class of bio-inspired continuous time RNNs with compact architecture for training the agent using expert demon strations collected using a classical controller. We propose a novel loss function that serves as a closed-loop feedback to the architecture, enabling the drone to dynamically adjust its actions and successfully navigate toward the target. This approach minimizes the need for extensive data collection and improves the system’s performance on the existing dataset. Additionally, we analyze the model’s attention maps, revealing that it effectively focuses on the target during navigation. Our findings highlight a detailed data collection process and provide a framework for reducing data requirements while simulating and implementing this approach in controlled environments.