Bilinear Mamba-Koopman Neural MPC for Varying Dynamics
Researchers have developed a new Bilinear Mamba-Koopman Neural MPC model that enhances model-predictive control for systems with varying dynamics. This model introduces control-dependent coupling in latent dynamics, allowing for better adaptation to changing conditions within a single control horizon. Experiments on CartPole and RSCP benchmarks showed improved forecasting accuracy and stabilization, particularly in time-varying scenarios and under delayed re-planning. AI
IMPACT Introduces a novel control-dependent latent dynamics mechanism for adaptive MPC, potentially improving performance in dynamic environments.