New AI Makes Robots More Reliable When Sensors Glitch

A new paper proposes a control framework for nonlinear systems facing intermittent measurements, a challenge that creates prediction uncertainty and disrupts state updates. The approach combines a Koopman-based stochastic model predictive controller with a learned linear latent predictor, while modeling measurement dropouts as a two-mode Markov chain. Under verifiable conditions, the method ensures bounded prediction errors and uses a probabilistic error radius to tighten constraints. An exact-penalty soft-constraint mechanism handles resets and prolonged dropouts. The article establishes recursive feasibility and stability guarantees, and numerical simulations on a visual-servoing task show effective tracking despite stochastic measurement unavailability.