The nuPlan benchmark is the first real-world autonomous driving dataset and evaluation framework, designed to assess the performance of machine learning (ML)-based planners in handling diverse driving scenarios and making safe, efficient decisions. The dataset includes 1282 hours of driving data from four cities (Las Vegas, Boston, Pittsburgh, and Singapore), along with high-quality auto-labeled object tracks and traffic light data. It also features a simulation framework that enables closed-loop testing of planners, allowing them to interact with other traffic participants. The benchmark includes a detailed analysis of various planning methods, highlighting the gaps between ML-based and traditional planning approaches. The dataset and simulation framework are publicly available at nuplan.org. nuPlan provides a modular and flexible simulation framework that supports different datasets and setups. The simulation is initialized with real-world observations, and an agent model predicts future trajectories. These are then passed to a planner to determine the best route for the ego vehicle, which is converted into a feasible trajectory by a controller. The simulation can be in open-loop (following ground-truth trajectories) or closed-loop (allowing the planner to deviate from recorded actions). The dataset includes detailed human-annotated 2D high-definition semantic maps, which are useful for navigation and planning. Auto-labeling techniques are used to generate object tracks and traffic light statuses, enabling the simulation of realistic driving scenarios. The benchmark also includes a scenario taxonomy and mining algorithms to provide fine-grained insights into planning performance. The nuPlan challenge demonstrated that rule-based planners outperform purely ML-based ones, but hybrid planners with learned components show the most promise in handling difficult scenarios. The challenge results also indicate that ML-based methods require additional post-processing for closed-loop driving, and hybrid methods appear to be the most effective approach, combining traditional and data-driven methods. The dataset and evaluation framework are publicly available, and the benchmark aims to advance research in learning-based planning by providing a comprehensive testbed for open-loop and closed-loop planning in real-world scenarios. The nuPlan benchmark is a significant step forward in the development of autonomous driving systems, offering a valuable resource for researchers and developers in the field.The nuPlan benchmark is the first real-world autonomous driving dataset and evaluation framework, designed to assess the performance of machine learning (ML)-based planners in handling diverse driving scenarios and making safe, efficient decisions. The dataset includes 1282 hours of driving data from four cities (Las Vegas, Boston, Pittsburgh, and Singapore), along with high-quality auto-labeled object tracks and traffic light data. It also features a simulation framework that enables closed-loop testing of planners, allowing them to interact with other traffic participants. The benchmark includes a detailed analysis of various planning methods, highlighting the gaps between ML-based and traditional planning approaches. The dataset and simulation framework are publicly available at nuplan.org. nuPlan provides a modular and flexible simulation framework that supports different datasets and setups. The simulation is initialized with real-world observations, and an agent model predicts future trajectories. These are then passed to a planner to determine the best route for the ego vehicle, which is converted into a feasible trajectory by a controller. The simulation can be in open-loop (following ground-truth trajectories) or closed-loop (allowing the planner to deviate from recorded actions). The dataset includes detailed human-annotated 2D high-definition semantic maps, which are useful for navigation and planning. Auto-labeling techniques are used to generate object tracks and traffic light statuses, enabling the simulation of realistic driving scenarios. The benchmark also includes a scenario taxonomy and mining algorithms to provide fine-grained insights into planning performance. The nuPlan challenge demonstrated that rule-based planners outperform purely ML-based ones, but hybrid planners with learned components show the most promise in handling difficult scenarios. The challenge results also indicate that ML-based methods require additional post-processing for closed-loop driving, and hybrid methods appear to be the most effective approach, combining traditional and data-driven methods. The dataset and evaluation framework are publicly available, and the benchmark aims to advance research in learning-based planning by providing a comprehensive testbed for open-loop and closed-loop planning in real-world scenarios. The nuPlan benchmark is a significant step forward in the development of autonomous driving systems, offering a valuable resource for researchers and developers in the field.