Mapless Navigation: Learning UAVs Motion forExploration of Unknown Environments

TL;DR

This paper introduces a reinforcement learning-based method enabling UAVs to navigate autonomously in unknown underground and indoor environments without relying on detailed maps, reducing computational costs while maintaining effective exploration.

Contribution

It proposes a novel mapless navigation approach using simulation-trained policies that transfer directly to real-world environments for autonomous UAV exploration.

Findings

Achieves mapless navigation with comparable performance to grid-based methods

Reduces computational cost of path planning

Successfully applied in underground mines and indoor spaces

Abstract

This study presents a new methodology for learning-based motion planning for autonomous exploration using aerial robots. Through the reinforcement learning method of learning through trial and error, the action policy is derived that can guide autonomous exploration of underground and tunnel environments. A new Markov decision process state is designed to learn the robot's action policy by using simulation only, and the results are applied to the real-world environment without further learning. Reduce the need for the precision map in grid-based path planner and achieve map-less navigation. The proposed method can have a path with less computing cost than the grid-based planner but has similar performance. The trained action policy is broadly evaluated in both simulation and field trials related to autonomous exploration of underground mines or indoor spaces.