Robustness as Viewpoint-Shift Magnitude Increases

Characterize how the magnitude of unseen camera-viewpoint shifts affects VersaCamVLA’s robustness and task success, particularly across different absolute azimuth-offset ranges.

Background

The paper evaluates VersaCamVLA under unseen camera poses and reports aggregate performance over a broad range of perturbations. However, camera-pose shifts can vary substantially in magnitude, and aggregate results do not establish how performance changes as the viewpoint displacement becomes larger.

The appendix identifies this unresolved issue and addresses it empirically by stratifying RoboTwin 2.0 evaluations according to absolute azimuth offsets of 0–30°, 30–60°, and 60–90°. Although the reported results suggest that VersaCamVLA remains stable across these evaluated intervals, the dependence of robustness on viewpoint-shift magnitude is explicitly identified as unresolved in the paper.

References

The unseen-pose results in Table~\ref{tab:camera_pose_ablation} average over a broad range of camera configurations, leaving the effect of viewpoint-shift magnitude unresolved.

— VersaCamVLA: Camera-Configurable VLA Policies for Robotic Manipulation  (2610.12451 - Han et al., 8 Oct 2026) in Appendix, Section “Generalization to Challenging Unseen Camera Pose Shifts” (Section \ref{app:large_pose_shifts})