What question did this study set out to answer?

To improve path planning and motion control for unmanned vehicles in dynamic environments through reinforcement learning.

March 15, 2026Open Access

Path Planning and Motion Control of Unmanned Vehicles Based on Reinforcement Learning

Key Points

To improve path planning and motion control for unmanned vehicles in dynamic environments through reinforcement learning.
Designed a collaborative architecture combining A* global planning with reinforcement learning for local control.
Utilized A* algorithm for generating global guidance paths.
Employed deep reinforcement learning for dynamic obstacle avoidance and trajectory tracking.
Conducted experiments in Gazebo simulation across three typical scenarios.
Achieved a navigation success rate of 96% in static obstacle environments.
Reduced path length by 6.2% compared to traditional methods.
Maintained a success rate above 92% in dynamic obstacle environments with a 17.4% increase in average collision distance.
Reduced task completion time by 15.3% in moving target point scenarios.

Abstract

Path planning and motion control are core technologies for autonomous navigation of unmanned vehicles (UAVs). Addressing the problems of poor adaptability and reliance on precise models in dynamic environments, traditional methods propose a reinforcement learning-based path planning and motion control method for UAVs, focusing on solving key technical issues in practical deployment from an application perspective. First, a collaborative architecture integrating A* global planning and reinforcement learning local control is constructed. The A* algorithm generates the global guidance path, while a deep reinforcement learning agent is responsible for dynamic obstacle avoidance and trajectory tracking. Based on this, the engineering implementation details, including state space design, reward function construction, and network structure configuration, are systematically described. Experiments in three typical scenarios are conducted in the Gazebo simulation environment. The results show that in static obstacle environments, the proposed method achieves a navigation success rate of 96%, with a path length reduction of 6.2% compared to traditional methods; in dynamic obstacle environments, the success rate remains above 92%, with an average collision distance increase of 17.4%; and in moving target point scenarios, task completion time is reduced by 15.3%. Finally, key issues and solutions in practical applications, such as Sim2Real transfer and multi-machine collaboration, are discussed.

AIに質問

Bookmark

View Full Paper