The paper introduces nuPlan, the world's first real-world autonomous driving dataset and benchmark designed to test the ability of machine learning (ML)-based planners to handle diverse driving situations and make safe, efficient decisions. The dataset includes 1282 hours of driving scenarios from four cities (Las Vegas, Boston, Pittsburgh, and Singapore), with high-quality auto-labeled object tracks and traffic light data. The authors also develop a closed-loop simulation and evaluation framework to enable planners' actions to be simulated in closed-loop, accounting for interactions with other traffic participants. The paper presents a detailed analysis of various baselines and investigates the gaps between ML-based and traditional methods. The nuPlan dataset and code are available at [nuplan.org](https://nuplan.org). The introduction highlights the challenges in generalizing driving scenarios from limited training data and the lack of formal safety guarantees in ML-based planning. The related work section discusses existing datasets, simulators, and planning methods, including classical, learning-based, and hybrid approaches. The dataset collection section describes the data collection process, auto-labeling techniques, and scenario mining. The simulation section details the modular and flexible simulation framework, including agent models, controllers, and evaluation metrics. The experiments section presents planning baselines and their results, demonstrating the impact of lower-quality perception inputs and generalization across different cities. The nuPlan challenge section highlights the performance of top planners in handling diverse scenarios. The conclusion emphasizes the significance of nuPlan as the largest existing labeled autonomous driving dataset and the need for further research in planning and sensor data integration.The paper introduces nuPlan, the world's first real-world autonomous driving dataset and benchmark designed to test the ability of machine learning (ML)-based planners to handle diverse driving situations and make safe, efficient decisions. The dataset includes 1282 hours of driving scenarios from four cities (Las Vegas, Boston, Pittsburgh, and Singapore), with high-quality auto-labeled object tracks and traffic light data. The authors also develop a closed-loop simulation and evaluation framework to enable planners' actions to be simulated in closed-loop, accounting for interactions with other traffic participants. The paper presents a detailed analysis of various baselines and investigates the gaps between ML-based and traditional methods. The nuPlan dataset and code are available at [nuplan.org](https://nuplan.org). The introduction highlights the challenges in generalizing driving scenarios from limited training data and the lack of formal safety guarantees in ML-based planning. The related work section discusses existing datasets, simulators, and planning methods, including classical, learning-based, and hybrid approaches. The dataset collection section describes the data collection process, auto-labeling techniques, and scenario mining. The simulation section details the modular and flexible simulation framework, including agent models, controllers, and evaluation metrics. The experiments section presents planning baselines and their results, demonstrating the impact of lower-quality perception inputs and generalization across different cities. The nuPlan challenge section highlights the performance of top planners in handling diverse scenarios. The conclusion emphasizes the significance of nuPlan as the largest existing labeled autonomous driving dataset and the need for further research in planning and sensor data integration.