[論文レビュー] Motion Planning for Autonomous Driving: The State of the Art and Future Perspectives

Siyu Teng, Xuemin Hu+9 more|arXiv (Cornell University)|Mar 17, 2023

Robotic Path Planning Algorithms被引用数 11

ひとこと要約

自動運転のモーションプランニングに関する総合的な調査で、パイプラインとエンドツーエンドのアプローチを対比し、課題・データセット・プラットフォーム・将来の方向性を概説する。

ABSTRACT

Intelligent vehicles (IVs) have gained worldwide attention due to their increased convenience, safety advantages, and potential commercial value. Despite predictions of commercial deployment by 2025, implementation remains limited to small-scale validation, with precise tracking controllers and motion planners being essential prerequisites for IVs. This paper reviews state-of-the-art motion planning methods for IVs, including pipeline planning and end-to-end planning methods. The study examines the selection, expansion, and optimization operations in a pipeline method, while it investigates training approaches and validation scenarios for driving tasks in end-to-end methods. Experimental platforms are reviewed to assist readers in choosing suitable training and validation strategies. A side-by-side comparison of the methods is provided to highlight their strengths and limitations, aiding system-level design choices. Current challenges and future perspectives are also discussed in this survey.

研究の動機と目的

• Classify and analyze motion planning methods for autonomous driving into pipeline and end-to-end frameworks.
• Examine expansion and optimization in pipeline planning and training, validation, and deployment in end-to-end planning.
• Survey experimental platforms, datasets, and validation scenarios to guide method selection.
• Provide a side-by-side comparison of strengths, limitations, and applicability for system design.
• Discuss current challenges and future directions to advance deployment of intelligent vehicles.

提案手法

• Review pipeline planning with global route planning and local trajectory planning, focusing on expansion and optimization mechanisms (state grid identification, primitive generation, etc.).
• Survey end-to-end planning methods categorized by learning approach: imitation learning, reinforcement learning, and parallel planning.
• Analyze training strategies, generalization, robustness, and deployment aspects of end-to-end models.
• Summarize datasets, simulation platforms, and physical platforms used for development and validation.
• Provide a side-by-side comparison of pipeline vs. end-to-end frameworks highlighting strengths and limitations.

実験結果

リサーチクエスチョン

• RQ1 What are the main distinctions, advantages, and drawbacks of pipeline versus end-to-end motion planning for autonomous driving?
• RQ2 How do current imitation learning, reinforcement learning, and parallel planning approaches perform and generalize across driving tasks and scenarios?
• RQ3 What datasets, simulation platforms, and real-world validation strategies support the development of IV motion planning methods?
• RQ4 What are the key challenges and future directions to enable reliable, safe, and scalable autonomous driving systems?

主な発見

• Pipeline planning offers interpretability and modularity but may suffer from suboptimal generalization and robustness due to hand-crafted components.
• End-to-end planning can improve generalization and robustness by learning task-specific representations but faces interpretability and validation challenges.
• A novel category called parallel planning is introduced to enhance end-to-end planning through virtual-real interaction learning.
• A wide range of datasets, simulation platforms, and semi-open real-world scenarios support advancing IV motion planning, though bridging sim-to-real gaps remains critical.
• Current challenges include perception reliability, safety against adversarial attacks, dataset-to-reality transfer, and governance considerations for deployment.