sparsedrive

SparseDrive is presented as an end-to-end autonomous driving stack that merges perception, prediction, and planning into a single differentiable pipeline, departing from the traditional modular setup where each task is handled separately. According to the NVIDIA NIM model card, the system encodes multi-view images into feature maps, learns a sparse scene representation through symmetric sparse perception, and then performs motion prediction and planning in parallel, producing a safe planning trajectory as output. The research paper describes SparseDrive as addressing the heavy compute cost of bird's-eye-view features by unifying detection, tracking, and online mapping into a fully sparse representation, paired with a parallel motion planner that models planning as a multi-modal problem and applies a hierarchical, collision-aware selection strategy to choose a rational trajectory.

The architecture is organized around two core innovations: a symmetric sparse perception module that consolidates detection, tracking, and online mapping with shared sparse features, and a parallel motion planner that treats prediction and planning jointly to improve safety and efficiency. This design is reported to surpass previous state-of-the-art end-to-end methods across perception, prediction, and planning benchmarks while delivering higher training and inference efficiency, making the approach attractive for production autonomous vehicle stacks. SparseDrive is hosted as a free endpoint on NVIDIA NIM under automotive and autonomous vehicle tags, with an accompanying API reference, positioning it as a ready-to-integrate building block for developers building full self-driving pipelines.