Dual-Attention Heterogeneous GNN for Multi-robot Collaborative Area Search via Deep Reinforcement Learning

TL;DR

The paper tackles multi-robot area search by addressing the exploration-coverage trade-off with a Dual-Attention Heterogeneous Graph Neural Network (DA-HGNN). It constructs a relational, spatio-temporal heterogeneous graph and augments it with type-aware semantics to distinguish frontier (exploration) from points of interest (coverage), enabling dynamic, task-aware long-term goal assignment via graph matching. The approach is embedded in a hierarchical policy, combining a reconstruction mapper, a topological planner, and an action planner, and trained with PPO plus an auxiliary perception loss. Empirical results in interactive 3D iGibson environments show DA-HGNN outperforms baselines in exploration, coverage, and scalability across varying scene sizes and robot counts, validating its generalization and practical impact for large-scale, dynamic multi-robot search tasks.

Abstract

In multi-robot collaborative area search, a key challenge is to dynamically balance the two objectives of exploring unknown areas and covering specific targets to be rescued. Existing methods are often constrained by homogeneous graph representations, thus failing to model and balance these distinct tasks. To address this problem, we propose a Dual-Attention Heterogeneous Graph Neural Network (DA-HGNN) trained using deep reinforcement learning. Our method constructs a heterogeneous graph that incorporates three entity types: robot nodes, frontier nodes, and interesting nodes, as well as their historical states. The dual-attention mechanism comprises the relational-aware attention and type-aware attention operations. The relational-aware attention captures the complex spatio-temporal relationships among robots and candidate goals. Building on this relational-aware heterogeneous graph, the type-aware attention separately computes the relevance between robots and each goal type (frontiers vs. points of interest), thereby decoupling the exploration and coverage from the unified tasks. Extensive experiments conducted in interactive 3D scenarios within the iGibson simulator, leveraging the Gibson and MatterPort3D datasets, validate the superior scalability and generalization capability of the proposed approach.

Figures (5)

• Figure 1: The overall framework of our method. The multi-robot area search method comprises three modules: a Reconstruction mapper, a Topological planner, and an Action planner. The topological planner is responsible for assigning a long-term goal to each robot over a heterogeneous graph. This is achieved through a structured three-stage process: Relational-aware graph construction, Type-aware graph augmentation, and Similarity measurement.
• Figure 2: Illustration of node feature update operation in the heterogeneous graph. The node features of the robots at layer $l$ are enhanced through interactions with the other entities in the cross-graphs.
• Figure 3: The scenes of the Gibson and Matterport3D datasets, which are collected from real indoor spaces. The robots can obtain egocentric RGBD images and their corresponding positions in a simulation scene.
• Figure 4: Performance of the testing scenes for exploration and coverage, compared with baseline methods. (a) Exploration percentage. (b) Coverage percentage. Both metrics validate the effectiveness of sub-tasks: the exploration percentage demonstrates the exploration capability (shown in (a)), while the coverage percentage illustrates the ability in coverage (shown in (b)).
• Figure 5: Visualization result in the simulation scene "5LpN3gDmAk7". We deploy three robots in the testing scene, and three colors denote the moving trajectories of the three robots.