Most diffusion-based planners for autonomous driving generate trajectories along a Gaussian-noise axis, treating driving utility as an afterthought. QualiDrive, proposed in a recent arXiv paper, instead formulates diffusion along a driving-quality axis. The method constructs scene-specific quality ladders by ranking observed trajectories according to measured driving utility, then learns to reverse this ordered degradation through low-to-high-quality refinement.
Each reverse step is conditioned on the scene, ego state, current trajectory, and target quality level, progressively refining trajectories toward higher driving quality. This reframes diffusion not as a denoising process but as a sequence of quality-ordered improvements.
On the NAVSIM benchmark, QualiDrive achieves 94.5 PDMS on v1 and a state-of-the-art 92.74 EPDMS on v2. The gain in Oracle EPDMS is modest (98.52 to 98.57), but the fractions of candidates above 0.95 and 0.98 EPDMS jump from 12.94%/9.71% to 21.53%/15.93%, raising selected EPDMS from 92.63 to 92.74. This suggests the improvement comes from a better distribution of candidate quality rather than a higher ceiling.
The authors argue that diffusion is not merely a process of denoising, but one of progressive quality-ordered refinement. As a single-source report, no conflicting findings are noted.