UniAD is an end-to-end autonomous driving framework that unifies perception, prediction, and planning tasks in a hierarchical, planning-oriented architecture.
The framework addresses the challenge of coordinating multiple autonomous driving subtasks by organizing them hierarchically rather than treating them as independent modules. Instead of multi-task learning with equal weight on all objectives, UniAD prioritizes planning as the top-level goal and structures perception and prediction as supporting tasks that feed into it. This planning-oriented philosophy means that intermediate representations and predictions are optimized specifically for their utility to the final planning decision, rather than for standalone accuracy.
UniAD suits teams building end-to-end autonomous driving systems who want a unified codebase rather than assembling separate perception, prediction, and planning modules. The framework is particularly valuable for researchers working with nuScenes data, though it is expanding to support nuPlan and NAVSIM datasets. The tool achieves state-of-the-art results across its component tasks, with motion prediction at 0.71m minADE, occupancy prediction at 63.4% IoU, and planning collision rates at 0.31% average collision. The framework has migrated to modern dependencies including mmdet3d 1.x and PyTorch 2.x, with existing workflows preserved for reproducibility.
The project maintains active development with recent framework upgrades to align with current library versions and expanded dataset support. Tools for nuPlan and NAVSIM benchmarking remain under development. The codebase includes comprehensive documentation covering installation, dataset preparation, training, and evaluation, with GPU requirements and example workflows documented for reproducibility.
