Learning-Based Autonomy from Kernel-Embedded Multi-modal Fusion to Feedback Control

ICLR 2026 Conference Submission14958 Authors

Keywords: Learning based autonomy, Kernel Embedded LinearModels, Deep Kernel Learning Bayesian Neural Networks

Abstract: In this work, we develop an end-to-end autonomy loop that couples \emph{kernel-embedded} multi-modal fusion with data-driven dynamics learning and feedback control. Heterogeneous sensor streams are embedded into a joint Reproducing Kernel Hilbert Space (RKHS) via additive/product kernels and conditional mean embeddings; dynamics are learned with kernel ridge regression (KRR), Deep Kernel Learning (DKL), or Bayesian deep neural networks (BDNNs); and policies are synthesized via dynamic programming (discrete and continuous-time HJB) or reinforcement learning with RKHS value functions. We present closed-form estimators, finite-sample and iteration-complexity characterizations, risk-sensitive planning with uncertainty, and safety via control barrier functions. We provide deployable algorithms, results and experiment in simulated robotics and precision irrigation.

Primary Area: learning on time series and dynamical systems

Submission Number: 14958