Candidate-based end-to-end driving planners generate a finite set of trajectory hypotheses and then select among them. But when the candidate set is sparse, it may not contain trajectories that fit the specific scene, leaving the planner without a good option. The abstract of a new arXiv paper highlights this problem, noting that a limited candidate set can provide insufficient support for scene-relevant maneuvers.

The proposed solution, called Expand-and-Prune, performs test-time candidate-space adaptation. As the name suggests, the method likely expands the candidate set to include more relevant trajectories and then prunes it to keep the search manageable. The abstract is brief, so details on how the expansion and pruning are implemented are not provided in the available text.

The work addresses a practical gap in sparse planning: fixed candidate sets may be too rigid for diverse driving contexts. By adapting the candidate space at test time, the method aims to give the planner better coverage of maneuvers without requiring a larger set at all times. Further details on the approach and its evaluation would be needed to assess its effectiveness.