CLOVER: Closed-Loop Value Estimation for Autonomous Driving Planning

To tackle the training-evaluation discrepancy in end-to-end autonomous driving planners, researchers have introduced CLOVER, a Closed-LOop Value Estimation and Ranking framework. Conventional planners rely on imitating a singular logged trajectory for training but assess performance using rule-based metrics such as safety, feasibility, progress, and comfort. This creates a situation where trajectories that closely resemble the logged path may breach rules, while those that deviate can still be valid and score highly. CLOVER employs a streamlined generator-scorer approach: the generator creates a variety of candidate trajectories, and the scorer estimates planning-metric sub-scores for ranking during inference. This enhances proposal support beyond mere single-trajectory imitation. The framework also develops an evaluator to boost candidate-set coverage and scorer ranking quality, which is crucial for proposal-selection planners. The research is published on arXiv (2605.15120).