SSAG: Situational Semantic Augmented Graph for Active SLAM in Object-Goal Navigation
Shasha Tian1,2, Zhengyang Chen1,3, Kai Ren1,2, Na Li1,2, Chongwei Ruan4, Zhijia Cui1,3, Mian Wu4,*
1 School of Computer Science, South-Central Minzu University, Wuhan, China
2 Hubei Provincial Engineering Research Center of Agricultural Blockchain and Intelligent Management, Wuhan, China
3 Hubei Provincial Engineering Research Center for Intelligent Management of Manufacturing Enterprises, Wuhan, China
4 DONG FENG Machine Tool PLANT Co., Ltd., Shiyan, China
* Corresponding Author: Mian Wu. Email: 
(This article belongs to the Special Issue: The Next-generation Deep Learning Approaches to Emerging Real-world Applications, 2nd Edition)
Computers, Materials & Continua https://doi.org/10.32604/cmc.2026.081556
Received 04 March 2026; Accepted 03 May 2026; Published online 15 June 2026
Abstract
To address the issues of low exploration efficiency and “geometric myopia” caused by the lack of high-level environmental structure modeling for mobile robots in complex indoor environments, this paper proposes an active SLAM object navigation method based on Situational Semantic Augmented Graph (SSAG). Unlike methods that learn policies solely on pixel-level semantic maps or exploit only object-level relations for implicit association, this work elevates local observations online into a room-level topological graph and performs explicit semantic reasoning over unobserved regions. First, an online room segmentation algorithm is employed to transform unstructured sensory data into a structured graph representation that characterizes room-level functional attributes and topological associations. Subsequently, a Graph Attention Network (GAT) is utilized to perform explicit reasoning on the semantic attributes of unobserved regions, providing decision-making priors for long-range exploration. Building upon this, we introduce a Product of Experts (PoE) mechanism within the Proximal Policy Optimization (PPO) framework to deeply fuse semantic reasoning heatmaps with geometric hard constraints. This integration optimizes the target selection strategy, generating long-term navigation goals with superior semantic consistency and physical reachability. Experimental results on the Habitat simulator and Gibson dataset demonstrate that the proposed SSAG achieves a Success weighted by Path Length (SPL) of 0.340, outperforming SemExp and SemGO by 15.3% and 4.9%, respectively, with a total success rate of 68.4%. Notably, for object search tasks with strong spatial correlations (e.g., “bed” and “toilet”), the success rates reach 73.2% and 72.5%, representing a performance gain of over 20% compared to baseline methods. These results validate the effectiveness of room-level semantic reasoning for object navigation and demonstrate the capability of the proposed method to achieve efficient autonomous navigation in unknown, complex scenarios.
Keywords
Active SLAM; object navigation; situational semantic augmented graph; graph attention network; deep reinforcement learning
Open Access
Login
Register
Submit a Paper
Propose a Special lssue
Download PDF
Downloads
Citation Tools
