Papers

Markovian Dynamics Enforcer: Feasibility Preserving Correction on Learned Dynamics Manifolds

This is a preprint that introduces MaDE, a frozen post-hoc operator that projects arbitrary neural trajectory predictions onto a learned dynamics manifold while inferring unobserved controls and reducing inequality violations. On inD intersection data, it cuts the one-step kinematic-bicycle residual from about 0.17 for raw predictors to 0.007, and cuts combined inequality violations from 3.6–5.8% to 0.7–1.1%, at the cost of raising average displacement error by 1.57–1.83x. This matters because it gives robot and vehicle pipelines a plug-and-play way to make data-driven forecasts dynamically and constraint consistent without retraining predictors.