Modernising Reinforcement Learning-Based Navigation for Embodied Semantic Scene Graph Generation

Summary

This post introduces an arXiv paper (2603.25415) on modernising reinforcement learning-based navigation for embodied semantic scene graph (SSG) generation. Semantic world models let embodied agents reason about objects, relations, and spatial context. The authors—Roman Kueble, Marco Hueller, Mrunmai Phatak, Rainer Lienhart, and Joerg Haehner (published 2026-03-26)—propose a modular navigation component for embodied SSG generation and modernise its decision-making by replacing policy optimisation methods and re-examining the discrete action formulation. They study compact versus larger, fine-grained discrete motion sets and compare a single-head policy over atomic actions with a decomposed multi-head policy over action components. Results show that merely swapping the optimisation algorithm improves SSG completeness by 21% relative to the baseline under identical reward shaping. Combining modern optimisation with fine-grained, decomposed action representations yields the strongest overall completeness and efficiency trade-off.