Multi-objective crowd-aware robot navigation system using deep reinforcement learning

Chien Lun Cheng, Chen Chien Hsu*, Saeed Saeedvand, Jun Hyung Jo

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review

Abstract

Navigating efficiently and safely through human crowds is essential for mobile robots in diverse applications such as delivery services, home assistance, healthcare, and manufacturing. However, traditional navigation methods are adversely affected by the high randomness of human movements, seriously hindering robot navigation in crowd environments. To tackle these problems, this paper proposes a deep reinforcement learning-based multi-objective crowd-aware robot navigation system called Multi-Objective Dual-Selection Reinforcement Learning (MODSRL). To deal with multiple objectives, including safety, time efficiency, collision avoidance, and path smoothness during navigation, a set of reward functions is used in MODSRL. To address the challenge of hesitation at the beginning when navigating in a crowd environment, a Dual-Selection Attention Module is developed, which enables the robot to make efficient decisions while reducing hesitation. Experimental results demonstrate that the proposed MODSRL outperforms existing approaches in terms of five different metrics. In particular, the average success rate of the proposed MODSRL outperforms ERVO, CADRL, LSTM-RL, and OM-SARL algorithms by 17.1%, 26.4%, 16.6%, and 9.2%, respectively, sufficing to show its robustness in complex crowd environments.

Original languageEnglish
Article number111154
JournalApplied Soft Computing
Volume151
DOIs
Publication statusPublished - 2024 Jan

Keywords

  • Deep Reinforcement Learning
  • Human Aware Motion Planning
  • Human-Robot Interaction
  • Obstacle Avoidance

ASJC Scopus subject areas

  • Software

Fingerprint

Dive into the research topics of 'Multi-objective crowd-aware robot navigation system using deep reinforcement learning'. Together they form a unique fingerprint.

Cite this