Collision Avoidance for Multiple UAVs in Unknown Scenarios with Causal Representation Disentanglement

TL;DR

The paper tackles the generalization gap in DRL-based multi-UAV collision avoidance caused by non-causal visual factors. It introduces causal representation disentanglement with a background intervention to separate task-relevant causal factors $S$ from non-causal factors $U$, feeding only $Z_2$ and $Z_3$ to the policy. The approach combines a VAE-based representation with losses $\\mathcal{L}_{vae}$, $\mathcal{L}_{rec}$, and $\mathcal{L}_C$, plus Fourier-domain background perturbations to create multi-domain data. Experiments in Unreal Engine/AirSim show robust generalization to unseen environments (grassland, snow mountain, forest) and initialization patterns, with clear improvements in Success Rate and SPL over the SAC+RAE baseline, demonstrating practical impact for reliable UAV operation in unknown scenarios.

Abstract

Deep reinforcement learning (DRL) has achieved remarkable progress in online path planning tasks for multi-UAV systems. However, existing DRL-based methods often suffer from performance degradation when tackling unseen scenarios, since the non-causal factors in visual representations adversely affect policy learning. To address this issue, we propose a novel representation learning approach, \ie, causal representation disentanglement, which can identify the causal and non-causal factors in representations. After that, we only pass causal factors for subsequent policy learning and thus explicitly eliminate the influence of non-causal factors, which effectively improves the generalization ability of DRL models. Experimental results show that our proposed method can achieve robust navigation performance and effective collision avoidance especially in unseen scenarios, which significantly outperforms existing SOTA algorithms.

Figures (6)

• Figure 1: The illustration of generalization ability analysis. When facing unseen scenarios, the current DRL method will suffer severe performance degradation in navigation success rate.
• Figure 2: Structural causal model (SCM) for representation learning. Image X consists of causal factors S and non-causal factors U, only causal factors S has a causal impact on the representation learning process.
• Figure 3: The architecture of our framework for Multi-UAV collision avoidance. The framework follows the SAC paradigm, which takes depth images, current velocity, and relative goal position as input and outputs flight control actions. We propose a causal representation disentanglement method to optimize the visual representation extraction and only pass the causal components for subsequent policy learning.
• Figure 4: The illustration of causal representation disentanglement. We use a variational auto-encoder for representation extraction, and then design specific loss function to guide pre-defined components to learn different semantic concepts. 'Reparam' represents the reparametrisation trick.
• Figure 5: Simulation scenarios for model training and testing. Specifically, playground scenario is used for model training, while grassland, snow mountain and forest scenarios are used for testing.