DEEP REINFORCEMENT LEARNING FOR MULTI-OBJECTIVE OPTIMISATION OF ROBOT ACTION SELECTION AND DYNAMIC OBSTACLE AVOIDANCE

Yixuan Zhu

Keywords

Deep reinforcement learning, dynamic obstacle avoidance, robotaction decision-making, state modelling, trajectory optimisation∗ School of Artificial Intelligence, Nanjing University of InformationScience and Technology, Nanjing 210044, Jiangsu, China; e-mail:[email protected] author: Yixuan ZhuRecommended by Shichao Liu

Abstract

To address the real-time and robustness requirements of robot navigation in dynamic environments, this paper proposes a deep reinforcement learning framework for multi-objective robot action decision and dynamic obstacle avoidance. A proximal policy optimisation (PPO)-based policy network is constructed with a weighted multi-objective reward function that jointly optimises collision risk, path efficiency, and trajectory smoothness. The state representation integrates LiDAR and inertial measurements to perceive spatial information and estimate the motion states of both static and dynamic obstacles, while the continuous action space outputs linear and angular velocity commands for fine-grained motion control. In addition, a Kalman filter- based short-term trajectory prediction module is incorporated into the policy layer, and a safety fallback mechanism is triggered when the predicted collision probability exceeds a predefined threshold, improving navigation reliability under rapidly changing conditions. The proposed method is evaluated in multi-scenario ROS–Gazebo simulations and compared with DWA, vanilla PPO, and representative DRL navigation baselines without prediction mechanisms. In high-density dynamic environments, the method achieves a collision rate of 2.3%, a target arrival rate of 96.7%, a 12.5% reduction in average path length, and a decision latency below 112 ms. The results demonstrate improved safety and navigation efficiency under simulated disturbance conditions, validating the effectiveness of integrating weighted reward optimisation, trajectory prediction, and continuous control for simulation-based mobile robot navigation.

Important Links:

Go Back