Task-Parameterized Dynamic Movement Primitives With Reinforcement Learning for Improved Motion Planning

Abstract

Online trajectory planning in unstructured environments poses significant challenges for mobile robots, particularly when navigating complex obstacles. Traditional learning-from-demonstration (LfD) methods depend on offline datasets, limiting their ability to adapt to varying obstacle shapes and dynamic conditions. To address these limitations, we propose a novel motion planning framework that combines global trajectory generation with local adaptability. Dynamic Movement Primitives (DMPs) are employed to generate global trajectories based on demonstrations, while Task-Parameterized Potential Fields (TPPFs) enhance local adaptability. The Policy Improvement through Path Integrals (PI<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$^{2}$</tex-math></inline-formula>) algorithm is utilized to optimize model parameters. The TPPF framework consists of two key components: (a) an obstacle avoidance field, which accounts for the robot's size, obstacle dimensions, and relative distances, allowing effective volumetric avoidance without extensive modeling; and (b) an attractive field, which directs the robot toward task-specific goals while steering it away from undesirable paths. By leveraging the PI<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$^{2}$</tex-math></inline-formula> algorithm, model parameters are optimized to produce trajectories that preserve the characteristics of demonstrated motions, while improving obstacle avoidance and task-oriented navigation. Experiments conducted in both simulations and dynamic real-world scenarios validate the proposed framework's effectiveness, demonstrating smoother trajectories and enhanced obstacle avoidance compared to baseline approaches.