Robustness under shifted observation histories and reactive traffic

Evaluate the robustness of SIMVTA under shifted observation histories and reactive traffic to determine whether its simulation-grounded video–trajectory alignment remains effective beyond log-based observations and non-reactive traffic settings.

Background

SIMVTA is trained using log-based observation histories and simulator-rendered targets associated with candidate trajectories. The authors note that current-policy trajectories broaden action coverage, while reconstructed targets may differ from real images and candidate filtering limits supervision. They also state that teacher-forced target selection does not capture autoregressive error accumulation.

The unresolved issue is whether the framework remains robust when observation histories are distributionally shifted and surrounding traffic is reactive, conditions that are more representative of challenging closed-loop deployment than the evaluated log-based and non-reactive settings.