WilliamXuanYu / CLOVER

What is CLOVER?

CLOVER is a research-backed planning framework for autonomous vehicles that bridges the gap between how driving models are trained and how they're evaluated. Most end-to-end planners learn by copying a single logged trajectory, but they're scored on rule-based metrics measuring safety, feasibility, and comfort. This creates a fundamental mismatch. CLOVER fixes this by generating diverse trajectory proposals, scoring them against real planning metrics, and refining the generator toward high-scoring options. Built in Python with PyTorch, it uses a DINOv2 vision backbone for perception and a transformer-based trajectory decoder. The system supports two training stages: first creating pseudo-expert trajectories that pass planning rules, then doing conservative self-distillation where a scorer guides generator refinement.

Why is it gaining traction?

The hook here is the training-evaluation alignment problem. If you've worked on motion planning, you know the frustration of models that look good on imitation losses but fail basic safety checks at runtime. CLOVER attacks this directly with a scorer that predicts sub-scores like collision avoidance, drivable area compliance, and time-to-collision. The multi-expert approach also stands out—instead of one "correct" trajectory, it trains on diverse high-scoring options, which should produce more robust behavior in edge cases. The preview training scripts let you experiment now, even if the official training pipeline is still in progress.

Who should use this?

Autonomous vehicle researchers working on planning and control who need better training signal alignment. Simulation engineers evaluating trajectory proposal methods will find the scorer architecture useful. This is not for production deployment yet—the code is academic, the star count is minimal, and several components remain unreleased. If you're evaluating closed-loop planning frameworks for research purposes, this is worth a look. If you need production-ready motion planning, look elsewhere.

Verdict

CLOVER addresses a real problem with a thoughtful two-stage approach, but the 1.0% credibility score reflects an early-stage research release. The preview training code works but carries stability warnings, and the official training pipeline, pseudo-expert generation, and NAVSIM-v2 evaluation scripts are still pending. Use it for research exploration and benchmarking, but expect to invest time in setup and handle some rough edges. The paper is solid; the engineering maturity needs patience.