arXiv Paper Introduces Safety-Critical Contextual Control Framework Using Riemannian Optimization

The paper "Safety-Critical Contextual Control via Online Riemannian Optimization with World Models," identified as arXiv:2604.19639v1, has been released on arXiv. This research tackles the difficulties of managing intricate systems lacking explicit dynamical descriptions. It introduces a sample-based Penalized Predictive Control framework that leverages online Riemannian optimization. In this method, a Planner refines task objectives by utilizing feasibility samples from a black-box Simulator, influenced by a context signal ξ_t. The Simulator transforms the feasibility manifold into a score-based density, which equips the action space with Riemannian geometry, aiding the Planner's gradient descent. A significant discovery reveals that the barrier curvature κ(ξ_t)—the lowest curvature of the conditional log-density—determines both the convergence rate and safety margin, supplanting the Lipschitz constant for unknown dynamics. The principal theoretical outcome presents a contextual safety bound that regulates the distance from the actual feasibility manifold, enhancing safety-critical control strategies in environments characterized by complex, implicit world models.