Autonomous Exploration of Mobile Robots in Complex Environments

復雜環境下的移動機器人自主探索

Student thesis: Doctoral Thesis

View graph of relations

Author(s)

Related Research Unit(s)

Detail(s)

Awarding Institution
Supervisors/Advisors
  • Lidai WANG (Supervisor)
  • Jia Pan (External person) (External Co-Supervisor)
Award date7 Dec 2023

Abstract

Autonomous exploration is crucial for mobile robots, but practical applications present a significant challenge due to dynamic obstacles. Moving pedestrians, in particular, poses a critical obstacle for deploying autonomous mobile robots in populated areas such as malls, airports, and museums. These dynamic objects create significant challenges for collision-free navigation and accurate localization or mapping, potentially jeopardizing the robot's safety and exploration performance. The robot may become trapped among pedestrians, preventing progress in scene coverage, or it may collide with humans and cause harm. Such collisions can disrupt the exploration process or lead to significant drift and artefacts in simultaneous localization and mapping (SLAM).

This thesis proposes new strategies for mobile robots to perform exploration tasks in human-populated environments. To ensure collision-free navigation and effective exploration, we propose a hierarchical planner that integrates local and global data. The approach also uses a reinforcement learning (RL)-based obstacle avoidance algorithm, enabling the robot to safely navigate through pedestrian crowds while following the exploration planner's path. The proposed system undergoes extensive evaluation in simulation environments, with results demonstrating superior performance compared to existing methods in both exploration efficiency and localization/mapping accuracy.

To achieve a more robust and secure mobile robotic exploration system, our proposed framework tightly couples a reinforcement-learned navigation controller with an exploration planner that is enhanced by a recovery planner. The navigation controller generates a value function that describes the distribution of crowds around the robot. The exploration and recovery planners utilize this value function to minimize human-robot interruptions. We evaluate the proposed exploration framework against several methods on indoor benchmarks with pedestrians, confirming its advantages in terms of exploration efficiency, navigation safety, and SLAM quality.

To enhance the efficiency of autonomous exploration and the robot's understanding of its environment, we further investigate the co-prediction of crowd flow and building layout. Firstly, we predict the movement of pedestrians with the Kalman Filter to infer a future crowd field. Then, we leverage a method to reconstruct the layout of an environment from a partial grid map using both geometric features and predicted crowd flow. This predicted layout provides more precise information gain estimations from candidate locations and enables early exploration termination when no further relevant area is expected to be discovered.

Overall, this thesis proposes methods for mobile robots to perform autonomous exploration in crowded environments. The scientific merits of this thesis include but are not limited to (i) an efficient and safe exploration system for mobile robots in crowded environments, (ii) a more robust and secure solution for mapping human-populated areas with active-recovery behaviour, and (iii) an advanced perception approach for understanding the surrounding environment.