Determine the effect of coarse-to-fine trajectory refinement on latency and prediction quality

Determine how the fixed-stack coarse-to-fine trajectory refinement design used by S²Planner affects inference latency and trajectory prediction quality.

Background

S²Planner generates candidate trajectories and refines them through a fixed stack of attention layers rather than using diffusion-based iterative denoising. Although coarse-to-fine refinement is a common strategy in vision and planning, the paper does not establish how this particular design affects the trade-off between computational latency and prediction quality. Resolving this issue requires empirical evaluation of both planning accuracy and runtime under controlled conditions.