An adaptive multi-robot exploration framework for lunar lava tubes: reinforcement learning-driven self-organized task allocation and local heuristic path planning

Lunar lava tubes are considered to be an ideal site for the establishment of a permanent lunar base because of their properties like natural radiation shielding and temperature stability, and the exploration of lava tubes is a key section in the development of lunar bases. Adopting an exploration robot to directly enter the lava tube caves is the most direct solution to address the internal exploration. As the traditional single-robot exploration faces challenges such as perception limitations, kinematic constraints, and low fault tolerance rate, it is necessary to promote the collaborative multi-robot exploration to achieve full-area coverage of the lava tube interior environment, redundant task execution, and adaptive environment interaction. We propose a multi-robot adaptive exploration framework, which employs Reinforcement Learning (RL) optimized self-organized task allocation and local heuristic path planning to improve the exploration efficiency and adaptability of the robotic system in the lava cave environment. Firstly, a dynamic RL-driven task allocation method is proposed to cluster the exploration area using self-organizing mapping (SOM) and dynamically adjust the task allocation according to robot power, sensor range and task complexity. The load allocation is continuously optimized during task execution by a proximal policy optimization strategy, and adaptive adjustment within the system is achieved when the environment changes. Secondly, a local heuristic-guided adaptive Conflict-Based Search (CBS) method is proposed to replace the low-level search of the traditional CBS with a local heuristic A* search that can be automatically adjusted based on the lava tube terrain features, which improves the computational efficiency while ensuring the optimal path. Finally, the framework was tested in a simulation environment. Results show that our framework outperforms the traditional methods in terms of coverage (12.2% improvement), computational efficiency (61.2% improvement), and energy consumption (32.7% improvement). Our framework effectively enhances the adaptive and collaborative capabilities of multi robot systems in unknown environments, and also provides an efficient solution for future scenarios such as planetary underground exploration.