Geometry-Preserving Inference for Multi-Agent Motion Models

Multi-agent motion models generate the behavior of surrounding road users to support the development and evaluation of autonomous driving systems. Faster execution makes repeated simulation less costly, but deploying a learned simulator must preserve the behavior it was trained to produce. We study this requirement in SMART and its CAT-K and RLFTSim variants. We show that exporting these models can change the probabilities of simulated motions even when sampled trajectories match. The discrepancies arise from changes in angle calculations and rounding when selecting nearby agents and map elements. Correcting these operations preserves every sampled action across 128 evaluated scenes and produces nearly identical motion probabilities. A controlled GPU benchmark shows a 1.34–1.49× inference speedup with the model operations held fixed. These results show why faster simulation must be checked against the probabilities that generate traffic. Matching sampled trajectories alone can conceal changes to the behavior a simulator represents.