Robotics SINGLE SOURCE

QualiDrive Orders Driving Trajectories by Quality, Not Only by Noise

First reported 20 Oct, 04:00 UTCSource: arXiv Robotics (cs.RO)Last editorial activity 20 Oct, 04:00 UTCWrite a comment ↓

A self-driving car navigates through a bustling city street in San Francisco, capturing urban mobility in action. — Illustrative photo.Photo by Abhishek Navlakha on Pexels

A preprint listed on arXiv Robotics (cs.RO) proposes QualiDrive, a diffusion-based planner for end-to-end autonomous driving that arranges trajectory generation along a driving-quality axis. The preprint reports state-of-the-art results on the NAVSIM v2 benchmark.

How the quality ladder is built

According to the arXiv listing, current diffusion planners for autonomous driving use driving utility only as an auxiliary guidance or selection signal. QualiDrive instead assembles scene-specific quality ladders by sorting observed trajectories according to their measured driving utility.

The model is then trained to undo this ordered degradation, stepping from low to high quality. Each reverse step draws on the scene, the state of the ego vehicle, the current trajectory and a target quality level, so trajectories are refined step by step toward better driving quality.

Benchmark results on NAVSIM

The preprint reports 94.5 PDMS on the NAVSIM v1 benchmark and 92.74 EPDMS on v2, which it presents as state of the art.

Compared with supervision that scores only the endpoint, the authors describe the gain in Oracle EPDMS as modest, from 98.52 to 98.57. The distribution of candidate trajectories shifts more clearly: the share above 0.95 EPDMS rises from 12.94% to 21.53%, and the share above 0.98 EPDMS from 9.71% to 15.93%. Selected EPDMS climbs from 92.63 to 92.74.

The authors' interpretation

From these results the authors conclude that diffusion is more than denoising, and also involves gradual refinement ordered by quality. This reading is the authors' own and has not been independently tested in the material available.

Open questions

Does the ordered refinement hold on benchmarks and real-world driving data beyond NAVSIM?
How large are the gains in selected EPDMS compared with other recent planners?

Key facts

A preprint listed on arXiv Robotics (cs.RO) proposes QualiDrive.

QualiDrive builds scene-specific quality ladders by ranking observed trajectories by measured driving utility.

The method is trained to reverse an ordered degradation from low to high quality.

Each reverse step takes into account the scene, the ego state, the current trajectory and a target quality level.

On NAVSIM, QualiDrive reaches 94.5 PDMS on v1 and 92.74 EPDMS on v2, a state-of-the-art result on v2.