williamxuanyu/clover

CLOVER is a research codebase for training self-driving car planning systems more effectively. It comes from a paper published on arXiv in May 2026. The problem it addresses is a gap between how these planning systems are trained and how they are evaluated. Typical training works by having the AI copy a single recorded driving route: it learns by watching one example of a human or expert driver handling a situation. But the metrics used to judge whether a planner is good cover things like safety, whether the vehicle makes forward progress, whether it stays comfortable, and whether it avoids dangerous situations. These metrics do not just ask "did you copy the example exactly?" which means a route that closely mimics the recorded example can still score poorly, while a different route that diverges from the recording can actually score better. CLOVER addresses this by expanding training beyond copying a single trajectory. First it generates many candidate trajectories and filters them using the evaluation metrics, building a richer set of high-quality training examples beyond the single recorded path. Then it trains a scoring model that learns to predict how well each candidate will actually score. That scorer is used to guide a second round of training, where the planner is refined toward the candidates the scorer rates highest. The process alternates between refining the scorer and refining the planner. The repository provides code for running inference with a pretrained model, preview training scripts, and evaluation scripts for the NAVSIM benchmark, which is a standard test environment for autonomous driving planning. The README notes that the official training code and the trajectory generation code have not yet been released at the time of writing. Pre-trained model checkpoints are available through the GitHub releases page. Installation requires Python 3.8, PyTorch, and a toolkit called nuplan-devkit that the project depends on for driving simulation infrastructure.