SPARC: Single-Pass Scaling for Motion Forecasting with Conformal Bayesian Last Layers

Abstract

Human motion forecasters are increasingly accurate and fast, but reliable deployment requires uncertainty estimates that are structured, calibrated, and efficient. SPARC introduces a Bayesian-conformal uncertainty layer for motion forecasting: a deterministic backbone predicts the future mean, while a conjugate Bayesian last layer converts feature leverage into an analytic horizon-wise epistemic scale. This scale augments structured trajectory covariance without Monte Carlo sampling, and split conformal calibration produces prediction tubes with finite-sample validity under exchangeability.