Intelligent multi-robot exploration in non-exposed spaces: methods and challenges

Liuchun Li, Bisheng Yang, Chi Chen, Zhengfei Yan, Shangzhe Sun, Yuhang Xu, Ang Jin, Leyi ZhaoArtif. Intell. Rev 2025 · DOI

Abstract

Autonomous exploration in "non-exposed spaces"—such as lunar lava tubes, deep mines, and disaster-stricken buildings—presents unique challenges including GNSS denial, perceptual degradation, and communication constraints. This paper reviews the shift from traditional rule-based navigation to advanced AI-driven strategies. It emphasizes the role of heterogeneous multi-robot systems (MRS) that leverage decentralized coordination to operate effectively in extreme, disconnected environments.

Introduction

The exploration of non-exposed spaces is critical for search and rescue, mining, and planetary science. Unlike typical outdoor environments, these spaces are characterized by three primary constraints: the absence of GPS (GNSS-denied), degraded perception due to dust or low light, and severe communication blockages caused by physical structures. While single-robot systems provide a baseline, multi-robot systems (MRS) offer the necessary redundancy and speed for time-critical missions. However, managing a fleet in such chaotic environments introduces a significant "coordination tax."

Methods and Research Landscape

The field has evolved through a structured taxonomy of technological advancements:

  • Environmental Perception: Transitioning toward Bird’s-Eye View (BEV) fusion and the use of event cameras to mitigate motion blur and sensor failure in high-dynamic-range scenes.
  • Path Planning: Moving from classic algorithms like A* to Deep Reinforcement Learning (DRL) and Graph Neural Networks (GNNs), which better handle complex 3D topologies and variable robot counts.
  • Coordination: Shifting from centralized to decentralized, "communication-aware" strategies where robots manage signal strength as a resource.
Keyword co-occurrence network showing research hotspots in multi-robot exploration.

The intensity of research in this domain has seen a dramatic increase over the last two decades. Analysis of publication data reveals a surge from fewer than 100 papers annually before 2010 to over 1,200 in 2024. This growth is mirrored by the increasing complexity and success of field validations, most notably the DARPA Subterranean Challenge, which demonstrated the viability of heterogeneous teams in mapping kilometers of underground tunnels autonomously.

Growth in publications and citations related to multi-robot autonomous exploration (2004–2024).

Discussion and Future Frontiers

The most successful exploration strategies utilize heterogeneous teams, combining the speed of UAVs with the endurance and payload capacity of UGVs or quadrupeds. Despite these gains, several bottlenecks remain:

  • Sim-to-Real Gap: AI models often struggle with real-world sensor noise, such as LiDAR reflections in dusty environments.
  • Computational Constraints: Advanced models like Transformers are often too power-hungry for the edge hardware found on small drones.
  • Asynchronous Collaboration: Future systems must improve coordination capabilities for robots that spend long periods disconnected and only "sync up" maps when they cross paths.

Emerging technologies like Neural Radiance Fields (NeRFs) and robotics-specific foundation models are expected to redefine how robots represent and understand 3D spaces in the near future.

Summary updated

BibTeX
@article{li2025intelligent,
  author = {Liuchun Li and Bisheng Yang and Chi Chen and Zhengfei Yan and Shangzhe Sun and Yuhang Xu and Ang Jin and Leyi Zhao},
  title = {{Intelligent multi-robot exploration in non-exposed spaces: methods and challenges}},
  journal = {Artif. Intell. Rev.},
  year = {2025},
  doi = {10.1007/s10462-025-11395-4},
}
Figure