Cognitive Embodied Learning for Anomaly Active Target Tracking

The primary challenge in active object tracking (AOT) lies in maintaining robust and accurate tracking performance in the complex physical scenarios. Existing end-to-end frameworks based on deep learning and reinforcement learning often struggle with high computational costs, data dependency, and limited generalization, hindering their performance in practical applications. Although embodied intelligence (EI) is promising to enable agents to learn from physical interactions, it cannot tackle severe anomalies happened in the complex scenarios. In order to address this issue, we propose a novel embodied learning method, called the Cognitive Embodied Learning (CEL), which is inspired by the dual decisionmaking system of the human brain. The CEL can dynamically switch between normal tracking and anomaly handling modes, supported by specialized modules including the anomaly cognition module (ACM), the rule reasoning module (RRM), and the anomaly elimination module (AEM). Moreover, we further introduce the categorical objective function (COF) to address function non-measurability and data confusion caused by severe anomalies. Extensive unmanned aerial vehicle (UAV) anomaly active target tracking experiments in both simulated and real-world scenarios demonstrate the superior performance of our method. Compared to the state-of-the-art methods, the CEL achieves a 361.4% increase in the success rate and a 54.4% improvement of the task completion efficiency, which highlights the potential of CEL to advance the field of AOT and open new avenues for more robust and intelligent tracking systems in the challenging environments.