Determine the source of trajectory conservatism under VeSTA

Determine whether VeSTA achieves lower collision burden by using smaller local displacements, more rotations, or larger obstacle margins, using matched action-level geometric analysis.

Background

The appendix examines whether VeSTA’s collision reduction could be explained trivially by executing fewer actions or stopping earlier. The reported episode-length statistics do not support a universal-shortening explanation: VeSTA has longer median trajectories and longer successful trajectories on both R2R and RxR, while its lower mean length is concentrated in failed episodes.

Nevertheless, the available aggregate statistics do not reveal how VeSTA changes local motion. The authors explicitly identify the unresolved alternatives as smaller displacements, increased rotation, or larger obstacle-clearance margins, motivating action-level geometric analysis.

References

It rules out a universal no-motion explanation, but not every form of conservatism: without matched action-level geometry, the analysis cannot determine whether VeSTA uses smaller local displacements, more rotations, or larger obstacle margins.

— Revealing and Bridging the Physical Executability Gap in Vision-Language Navigation  (2610.25687 - He et al., 19 Oct 2026) in Appendix, Section A10, “Execution Length and Conservatism”