A Survey of Large Foundation Models for Trajectory Planning in Autonomous Driving

ABSTRACT Trajectory planning converts scene understanding, motion prediction, navigation objectives and traffic constraints into safe and executable ego motion. Its scope has expanded from geometric search and constrained optimization to representation learning oriented toward planning, end‐to‐end driving, semantic reasoning, generative trajectory modelling and action‐conditioned future prediction. This expansion has also blurred several distinct concepts: a planning paradigm describes how the driving stack is organized, whereas a planning method describes how trajectories are generated, evaluated, refined or verified. This survey develops a unified framework for organizing the field around the planning problem in the foundation model era. We first formulate trajectory planning as constrained decision making under uncertainty and distinguish open‐loop from closed‐loop operation and one‐stage from multi‐stage planning. We then connect structured, unstructured and open‐world scenarios to the data formats used to encode geometry, semantics, interaction, intent and future evolution. At the architectural level, we analyse modular, learning‐enhanced modular, planning‐oriented end‐to‐end and hybrid paradigms as overlapping rather than mutually exclusive research streams. At the methodological level, we organize the literature into classical autonomous driving trajectory planning algorithms, planning based on large language models and vision‐language models and planning based on world models, with detailed subcategories reflecting the functional role of each method. We further compare datasets, simulators and evaluation protocols across safety, task completion, comfort, rule compliance, planning quality, robustness, semantic grounding, world model utility and efficiency. Finally, we identify physical and semantic grounding, uncertainty propagation, closed‐loop generalization, real‐time generative inference, long‐tail robustness and verifiable hybrid integration as central challenges. The resulting taxonomy clarifies how established planning principles and foundation model capabilities can be combined without sacrificing executability or safety.

Authors

Institutions

Publication Details

Journal
Expert Systems
Published
2026-11-13
DOI
https://doi.org/10.1111/exsy.70482
Primary Topic
Autonomous Vehicle Technology and Safety
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

A Survey of Large Foundation Models for Trajectory Planning in Autonomous Driving

Chenchen Fu, Rufan Bai, Yuhuan Lu, Wei William Wang et al.
Expert Systems
Autonomous Vehicle Technology and Safety
article

A Survey of Large Foundation Models for Trajectory Planning in Autonomous Driving

Chenchen Fu, Rufan Bai, Yuhuan Lu, Wei William Wang, Tiantian Chen, Yixian Chen, Binbin Song, Tianyi Xue, Qiqi Chen, Tiantian Zhou, Jingjing Yang, Weiwei Wu, Jiaqi Li, Hui Han
article en

Abstract

ABSTRACT Trajectory planning converts scene understanding, motion prediction, navigation objectives and traffic constraints into safe and executable ego motion. Its scope has expanded from geometric search and constrained optimization to representation learning oriented toward planning, end‐to‐end driving, semantic reasoning, generative trajectory modelling and action‐conditioned future prediction. This expansion has also blurred several distinct concepts: a planning paradigm describes how the driving stack is organized, whereas a planning method describes how trajectories are generated, evaluated, refined or verified. This survey develops a unified framework for organizing the field around the planning problem in the foundation model era. We first formulate trajectory planning as constrained decision making under uncertainty and distinguish open‐loop from closed‐loop operation and one‐stage from multi‐stage planning. We then connect structured, unstructured and open‐world scenarios to the data formats used to encode geometry, semantics, interaction, intent and future evolution. At the architectural level, we analyse modular, learning‐enhanced modular, planning‐oriented end‐to‐end and hybrid paradigms as overlapping rather than mutually exclusive research streams. At the methodological level, we organize the literature into classical autonomous driving trajectory planning algorithms, planning based on large language models and vision‐language models and planning based on world models, with detailed subcategories reflecting the functional role of each method. We further compare datasets, simulators and evaluation protocols across safety, task completion, comfort, rule compliance, planning quality, robustness, semantic grounding, world model utility and efficiency. Finally, we identify physical and semantic grounding, uncertainty propagation, closed‐loop generalization, real‐time generative inference, long‐tail robustness and verifiable hybrid integration as central challenges. The resulting taxonomy clarifies how established planning principles and foundation model capabilities can be combined without sacrificing executability or safety.

Expert SystemsVol. 43(12)
Foshan University (CN), Korea Advanced Institute of Science and Technology (KR), University of Macau (MO), Harbin Institute of Technology (CN), Macao Polytechnic University (MO), Southeast University (CN)
Autonomous Vehicle Technology and Safety
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon