Structure-Graph Priors and Topology-Consistency Regularization for End-to-End Audio World Models: A Risk-Sensitive Protocol for Embodied Auditory Interaction

Abstract

Embodied robots increasingly rely on hearing for timely and safe responses in open environments. We propose an end-to-end Audio World Model (AWM) in a Dreamer/RSSM style with (i) a replaceable structure-graph prior on discrete cluster transitions, tested against random, wrong-mapping, and learned graphs, and (ii) topology-consistency regularization from frozen audio embeddings. Control uses risk-penalized, occupancy-constrained, or CVaR latent CEM planning; a reward head is trained on posteriors and evaluated on prior imagination to prevent leakage. This Version publishes methods and a frozen evaluation protocol; no measured closed-loop Success/DangerOcc/CVaR numbers are claimed. A small PyTorch helper (awm_ref.py) is included. Chinese and English full texts are deposited.