SparseDrive is a research codebase for camera-based autonomous driving perception and planning, published at ICRA 2025. It implements a unified end-to-end model that jointly performs 3D object detection, multi-object tracking, online HD map vectorization, motion prediction, and ego-motion planning using a sparse scene representation. The system is built on the mmdetection3d framework with custom plugins, and is trained and evaluated exclusively on the nuScenes dataset. The code includes full training pipelines (two-stage: perception first, then motion+planning), data conversion with extended annotations (future trajectories, drive commands, CAN bus ego status), custom CUDA kernels for deformable feature aggregation, and comprehensive evaluation metrics for all tasks. While not a production application, it is a well-structured research implementation that demonstrates non-trivial architectural innovations and provides tools for reproducibility.