UTracker: Learning Visuomotor Policies for Underwater Active Target Tracking via Imitation Learning and Diffusion Model

Abstract

Active visual tracking of underwater non-cooperative targets is a challenging task for autonomous underwater vehicles (AUVs) due to the complexity of underwater environments and the unpredictable dynamics of target motion. To address this challenge, this paper proposes UTracker, a novel framework for learning visuomotor policies tailored to underwater active target tracking. UTracker employs a two-stage training pipeline to enhance data efficiency and generalization. First, a state-based expert policy is trained in simulated underwater environments to perform underwater target tracking, and then used to generate paired image–action data as expert demonstrations. Second, a visuomotor policy is distilled from expert demonstrations via imitation learning, with diffusion models incorporated to refine action sequences and improve robustness against observation noise and dynamic perturbations. Extensive simulation studies show that UTracker outperforms baseline methods in challenging underwater scenarios. Moreover, sim-to-real experiments on a real AUV validate that the learned visuomotor policy reliably tracks non-cooperative targets in real-world environments. Video and code are available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/Ice-mao/RL_AUV_tracking.git</uri>.