Evaluate GS-VLA under broader rotation perturbations

Determine the performance of GS-VLA under rotation sweeps beyond the reported joint perturbation settings, while ensuring that sampled camera poses retain the workspace within the camera frustum.

Background

The experiments primarily vary translation and evaluate only four joint translation–yaw configurations. The authors explain that larger combined translation and rotation often moves the camera off the workspace, so many sampled views do not contain the workspace in their frusta. Consequently, the robustness of GS-VLA at larger rotational perturbations remains unevaluated, and the reported results do not establish its behavior beyond the tested regime.

References

We could not run a clean rotation sweep beyond the joint $(\varepsilon_t,\theta)$ cells reported in Table~\ref{tab:joint} for this reason.

— GS-VLA: Plug-and-Play Viewpoint Canonicalization for Frozen VLA Policies via Gaussian Splatting  (2608.19066 - Park et al., 19 Aug 2026) in Section 5, “Limitation,” subsection “Locality boundary on rotation”