Effect of explicit spatial structure on dynamics-model attention

Establish whether incorporating explicit spatial structure into learned dynamics models consistently improves their ability to attend to the most relevant environmental features, particularly for trajectories involving large displacements from the starting position.

Background

The paper introduces SpatialBEV, a recurrent dynamics architecture that queries localized bird’s-eye-view features along the vehicle’s predicted footprint rather than encoding the entire BEV representation as a fixed initial context. The authors report that the advantage of this structured representation is especially pronounced on the Speed dataset, where trajectories involve larger displacements from their starting positions.

The authors explicitly identify the claim that explicit spatial structure improves attention to relevant environmental features as a conjecture. Their reported Transformer comparisons support, but do not establish, this broader claim across architectures, datasets, and operating conditions.