SparseDrive is an end-to-end autonomous driving system that addresses the limitations of traditional modular systems by adopting a sparse scene representation approach. Traditional systems suffer from information loss and error accumulation across modules, while end-to-end paradigms unify tasks into a differentiable framework for optimization. However, existing methods face challenges in planning safety due to computationally expensive BEV features and simplistic task designs. SparseDrive introduces a new paradigm, SparseDrive, which combines a symmetric sparse perception module and a parallel motion planner. The perception module unifies detection, tracking, and online mapping using a symmetric architecture to learn sparse scene representations. The motion planner leverages the similarity between motion prediction and planning tasks, employing a parallel design and a hierarchical planning selection strategy with a collision-aware rescore module to ensure safety. SparseDrive achieves superior performance and efficiency, outperforming previous state-of-the-art methods in all tasks while significantly improving training and inference speed. The system is evaluated on the nuScenes benchmark, demonstrating high accuracy in perception, motion prediction, and planning, with a notable reduction in collision rate and L2 error. SparseDrive's design emphasizes sparse representation, multi-modal planning, and efficient training, making it a promising approach for end-to-end autonomous driving.SparseDrive is an end-to-end autonomous driving system that addresses the limitations of traditional modular systems by adopting a sparse scene representation approach. Traditional systems suffer from information loss and error accumulation across modules, while end-to-end paradigms unify tasks into a differentiable framework for optimization. However, existing methods face challenges in planning safety due to computationally expensive BEV features and simplistic task designs. SparseDrive introduces a new paradigm, SparseDrive, which combines a symmetric sparse perception module and a parallel motion planner. The perception module unifies detection, tracking, and online mapping using a symmetric architecture to learn sparse scene representations. The motion planner leverages the similarity between motion prediction and planning tasks, employing a parallel design and a hierarchical planning selection strategy with a collision-aware rescore module to ensure safety. SparseDrive achieves superior performance and efficiency, outperforming previous state-of-the-art methods in all tasks while significantly improving training and inference speed. The system is evaluated on the nuScenes benchmark, demonstrating high accuracy in perception, motion prediction, and planning, with a notable reduction in collision rate and L2 error. SparseDrive's design emphasizes sparse representation, multi-modal planning, and efficient training, making it a promising approach for end-to-end autonomous driving.