SFKD: New framework slashes autonomous vehicle tracking error by 31%
Stable Fiber-Koopman model ensures 44% smoother control and provable stability across changing environments.
Syed Pouladi's new paper proposes Stable Fiber-Koopman Residual Dynamics (SFKD), a framework that tackles the tension between model expressiveness and formal stability guarantees. SFKD constructs a fiber bundle latent manifold where each fiber encodes environment-specific dynamics (e.g., varying road friction or wind). An environment-conditioned Koopman operator governs the dominant linear evolution on each fiber, while a contraction-constrained residual neural network captures unmodeled nonlinear effects with an explicit input-to-state stability (ISS) certificate. This design ensures provable stability even as conditions shift.
In autonomous vehicle path tracking experiments with variable surface conditions and wind disturbances, SFKD outperformed five baselines including Koopman MPC and Neural ODE. It achieved a 31% reduction in tracking RMSE, a 44% improvement in control smoothness, and near-zero latent stability violation across environment-switching scenarios. The theoretical analysis establishes ISS of the latent dynamics and a finite ultimate bound on tracking error. This work offers a practical path to safer, more reliable learned controllers for robotics and autonomous systems.
- SFKD combines fiber bundle manifolds, environment-conditioned Koopman operators, and contractive residual NNs for provably stable learning-based control.
- Benchmarked against five methods (Koopman MPC, Neural ODE, ICODE, ControlSynth, ICODE-MPPI) with 31% lower RMSE and 44% smoother control.
- Theoretical ISS certificate ensures latent stability and bounded tracking error across environment switches.
Why It Matters
Enables safer autonomous systems by blending learned models with formal stability guarantees across changing environments.