ESCORT: Efficient Stein-variational and Sliced Consistency-Optimized Temporal Belief Representation for POMDPs

Traditional belief representation methods in POMDPs struggle with high-dimensional, multi-modal distributions due to kernel degeneracy and the need for excessive particles. Recent neural and parametric approaches fail to capture intricate correlation patterns essential for accurate decision-making. ESCORT is innovative because it combines principled geometric methods (sliced Wasserstein distance) with temporal consistency regularization, enabling particle-based methods to scale to complex belief spaces while preserving critical statistical dependencies that impact decision quality.

ESCORT is a particle-based framework for belief approximation in partially observable Markov decision processes that addresses the challenge of representing complex, multi-modal belief distributions in high-dimensional spaces. The approach extends Stein Variational Gradient Descent with correlation-aware projections and temporal consistency constraints, enabling particles to concentrate in high-uncertainty regions while preserving learned correlation structures. ESCORT dynamically adapts to belief landscape complexity without requiring resampling, maintaining both representational accuracy and computational efficiency.

Extensive evaluation on Light-Dark Navigation, Kidnapped Robot, and Multi-Target Tracking benchmarks demonstrates that ESCORT consistently outperforms state-of-the-art belief approximation methods including transformers and density-based approaches. The framework achieves superior belief fidelity and decision quality across domains ranging from discrete to continuous high-dimensional problems.