Local Lane Graph Conditioning as a General Inductive Bias for Trajectory Prediction: A Multi-Architecture Study on the Waymo Open Motion Dataset
TL;DR
This study isolates the effect of local lane graph conditioning by proposing a waterflow method that extracts an ego-centric lane topology via breadth-first traversal of the HD map, suggesting feasibility for resource-constrained deployment, pending validation on production automotive hardware.
Abstract
Accurate trajectory prediction is critical for autonomous driving safety and energy-efficient motion planning in sustainable urban mobility. This study isolates the effect of local lane graph conditioning by proposing a waterflow method that extracts an ego-centric lane topology via breadth-first traversal of the HD map (For convenience, all acronyms used throughout this manuscript (e.g., HD, LSTM, ADE, FDE, minADE, minFDE, BFS, GNN, SOTA) are collected in the Abbreviations section at the end of the paper), fusing lane features into trajectory encoders through cross-attention. The evaluation is deliberately scoped to ego-vehicle prediction at signal-controlled intersections: we evaluate across two architectures (LSTM and Transformer), two horizons (3 s and 8 s), and both single- and multi-modal (K=6) settings on 89,258 such scenarios from the Waymo Open Motion Dataset, so that “generality” refers to consistency across architectures, horizons, and output settings within this scope, rather than across prediction tasks. Lane conditioning consistently improves accuracy: +9.3% ADE at 3 s (p=0.007, 3 seeds), +26.6% minADE at 8 s (K=6, p=0.003, 3 seeds), and +26.8% ADE for the Transformer (p=0.030, 3 seeds)—with only ∼8% additional parameters for the LSTM. A controlled full-graph ablation (nearest 64 lanes) shows a consistent but not statistically significant trend favouring topologically guided local selection over brute-force spatial proximity (+11.4% minADE, p=0.063). Error decomposition reveals balanced lateral (+26.5%) and longitudinal (+25.4%) improvements, and a per-maneuver analysis over all 13,388 validation scenarios shows the largest gains for turning maneuvers. The lane-conditioned model (<700,000 parameters) runs in 0.8 ms per prediction on a desktop GPU (0.7 ms single-threaded CPU) with below 30 MB peak memory and an estimated 51 mJ per prediction, suggesting feasibility for resource-constrained deployment, pending validation on production automotive hardware.