FlowS: One-Step Motion Prediction via Local Transport Conditioning
Researchers achieve 0.48 mAP on Waymo with single-step inference, no denoising needed.
FlowS tackles a core tension in autonomous driving: generative motion prediction must be accurate, capture diverse multimodal futures, and run within strict latency budgets. Diffusion models excel at accuracy and diversity but require tens to hundreds of denoising steps, violating real-time constraints. FlowS resolves this by identifying that single-step integration works when the transport problem is local—meaning the model only needs to refine a trajectory already near a plausible future, not discover one from scratch.
FlowS introduces two key mechanisms. First, an online, scene-conditioned learned prior emits K calibrated anchor trajectories per agent, each near a plausible future. This converts mode discovery into local correction. Second, a step-consistent displacement field enforces semigroup self-consistency, so a single step inherits multi-step accuracy. Anchoring this field at learned priors along straight-line paths yields a stable, low-variance training target, unlike prior self-consistency methods that suffer from high-variance bootstrap signals on curved diffusion paths. On Waymo, FlowS achieves state-of-the-art results: Soft mAP 0.4804 and mAP 0.4703 with ensemble at 75 FPS.
- FlowS uses local transport conditioning to achieve single-step inference, avoiding the latency of diffusion models that require tens to hundreds of denoising steps.
- A scene-conditioned learned prior emits K calibrated anchor trajectories per agent, converting mode discovery into local correction.
- On Waymo Open Motion Dataset, FlowS achieves state-of-the-art Soft mAP (0.4804) and mAP (0.4703) with ensemble at 75 FPS.
- Step-consistent displacement field enforces semigroup self-consistency, ensuring single-step accuracy matches multi-step performance.
Why It Matters
FlowS makes one-step generative motion prediction practical for safety-critical autonomy, enabling real-time, accurate, and diverse trajectory forecasting.