Deep-reinforcement-learning-based multi-UAV collision avoidance and navigation methods have made significant progress. However, the fundamental challenge of those methods is their restricted capability to generalize beyond the specific scenario in which they are trained on. We find that the cause of the generalization failures is attributed to spurious correlation. To solve this generalization problem, we propose a causal representation learning method to identify the causal representations from images. Specifically, our method can neglect factors of variation that are irrelevant to the deep reinforcement learning task through causal intervention. Subsequently, the causal representations are fed into the policy network for action prediction. Extensive testing reveals that our proposed method exhibits better generalization results compared to state-of-the-art methods in different testing scenes.

Deep reinforcement learning (DRL) has significantly advanced online path planning for unmanned aerial vehicles (UAVs). Nonetheless, DRL-based methods often encounter reduced performance when dealing with unfamiliar scenarios. This decline is mainly caused by the presence of non-causal and domain-specific elements within visual representations, which negatively impact the learning of policies. Present techniques generally depend on predefined augmentation or regularization methods intended to direct the model toward identifying causal and domain-invariant components, thereby enhancing the model’s ability to generalize. However, these manually crafted approaches are intrinsically constrained in their coverage and do not consider the entire spectrum of possible scenarios, resulting in less effective performance in new environments. Unlike prior studies, this work prioritizes representation learning and presents a novel method for causal representation disentanglement. The approach successfully distinguishes between causal and non-causal elements in visual data. By concentrating on causal aspects during the policy learning phase, the impact of non-causal factors is minimized, thereby improving the generalizability of DRL models. Experimental results demonstrate that our technique achieves reliable navigation and effective collision avoidance in unseen scenarios, surpassing state-of-the-art deep reinforcement learning algorithms.