Motion Planning for Autonomous Driving: The State of the Art and Future Perspectives

Teng, Siyu; Hu, Xuemin; Deng, Peng; Li, Bai; Li, Yuchen; Ai, Yunfeng; Yang, Dongsheng; Li, Lingxi; Xuanyuan, Zhe; Zhu, Fenghua; Chen, Long · 2023 · OpenAlex-citations

DOI: 10.1109/tiv.2023.3274536

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper provides a comprehensive review of state-of-the-art motion planning methods for intelligent vehicles (IVs), addressing the critical need for reliable and safe autonomous driving systems. Despite the potential for commercial deployment, IV implementation remains limited due to challenges in reliability and safety. The authors categorize motion planning into two primary frameworks: pipeline planning, which is modular and rule-based, and end-to-end planning, which treats the entire driving task as a single learnable machine learning problem. The study aims to assist researchers and engineers in selecting appropriate methods by comparing their strengths, limitations, and underlying mechanisms. The review analyzes pipeline planning methods, which consist of global route planning and local behavior/trajectory planning. Global route planning utilizes algorithms like Dijkstra and A* to find optimal paths on road networks. Local planning is further decomposed into state grid identification, primitive generation, and other approaches. State grid identification employs search-based methods (e.g., Hybrid A*), selection-based methods (e.g., Markov decision processes), or optimization-based methods that discretize optimal control problems. Primitive generation connects these grids using closed-form rules, simulations, interpolation, or numerical optimization. The paper highlights that while pipeline methods offer high interpretability and ease of debugging, they suffer from high computational costs, reliance on manual heuristics, and potential robustness issues due to modular concatenation. Conversely, the paper examines end-to-end planning methods, which map raw sensor data directly to control commands without intermediate modules. This category is divided into imitation learning (IL), reinforcement learning (RL), and parallel learning. IL methods learn policies from expert trajectories using supervised learning, while RL develops policies through unsupervised learning and reward mechanisms. Parallel learning is introduced as a novel category involving virtual-real interaction confusion learning. The authors review specific implementations, including behavioral cloning and inverse reinforcement learning, noting that end-to-end methods offer superior generalization, robustness, and real-time capabilities by avoiding human-defined information bottlenecks. However, they face significant challenges in interpretability, making it difficult to trace the causes of model errors. The significance of this work lies in its structured comparison of these two paradigms and its inclusion of auxiliary elements such as datasets, simulation platforms, and physical testing scenarios. The authors propose a new classification for pipeline methods based on expansion and optimization mechanisms and introduce parallel planning as a distinct end-to-end category. By highlighting the trade-offs between the interpretability of pipeline methods and the performance potential of end-to-end methods, the paper provides essential insights for system-level design choices. It concludes by discussing current challenges and future perspectives, emphasizing the need for reliable validation strategies and improved interpretability in learning-based approaches to advance the field of autonomous driving.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• passenger motion sickness comfort