Data-Driven Motion Planning: A Survey on Deep Neural Networks, Reinforcement Learning, and Large Language Model Approaches

Abstract

Motion planning is a fundamental challenge in robotics, involving the creation of trajectories from start to goal states while meeting constraints like collision avoidance and joint limits. Its complexity increases with the number of robot joints. Several traditional approaches tackle this problem, such as sampling motion planning, grid-based methods, potential fields, and optimization techniques. Recent advancements in deep neural networks, reinforcement learning, and large language models enable new possibilities for solving motion planning problems by improving sampling efficiency, optimizing control policies, and enabling task planning through natural language prompts. This survey comprehensively reviews these novel approaches, providing background knowledge, analyzing key contributions, and identifying common patterns, limitations, and research gaps. Our work is the first to integrate all three major data-driven approaches, discussing their applications and future research directions.