RoAD Benchmark: How LiDAR Models Fail under Coupled Domain Shifts and Label Evolution

For 3D perception systems to operate reliably in real-world environments, they must remain robust to evolving sensor characteristics and changes in object taxonomies. However, existing adaptive learning paradigms struggle in LiDAR settings where domain shifts and label-space evolution occur simultaneously. We introduce $\mathbf{\text{Robust}}\mathbf{\text{Autonomous}}\mathbf{\text{Driving}}\mathbf{\text{under}}\mathbf{\text{Dataset}}\mathbf{\text{shifts}}\mathbf{\text{(RoAD)}}$, a benchmark for evaluating model robustness in LiDAR-based object classification under intertwined domain shifts and label evolution, including subclass refinement, unseen-class insertion, and label expansion. RoAD evaluates three learning scenarios with increasing adaptation, from fixed representations (zero-shot transfer and linear probing) to sequential updates (continual learning). Experiments span large-scale autonomous driving datasets, including Waymo, nuScenes, and Argoverse2. Our analysis identifies central failure modes: (i) $\mathit{\text{limited}}\mathit{\text{transferability}}$ under subclass refinement and unseen-class insertion, and on non-vehicle class; and (ii) $\mathit{\text{accelerated}}\mathit{\text{forgetting}}\mathit{\text{during}}\mathit{\text{continual}}\mathit{\text{adaptation}}$, driven by feature collapse and self-supervised learning objectives.