Adaptive parameter tuning for vSLAM: a graph-enhanced decision transformer approach

Abstract Visual simultaneous localization and mapping (vSLAM) is crucial for navigation and localization tasks in mobile robots and augmented/virtual reality. However, existing vSLAM methods typically rely on static parameter configurations, requiring manual tuning based on experience and specific application scenarios. This limits their adaptability and versatility. To address this challenge, we propose an adaptive parameter tuning method based on the graph-enhanced decision transformer(GEDT), an extension of the decision transformer, which transforms the parameter adaptation problem into a sequence modeling task and optimizes vSLAM parameters using a Gaussian Reward-Weighted Causal Graph (GRW-CG) structure. Specifically, we model the state, action, and reward within the adaptive parameter tuning process as a GRW-CG, where the state is represented as the source node, while actions are represented as nodes connected by edges. The edge weights are defined by a Gaussian distribution centered around the reward of the current state-action pair, thereby emphasizing the strong correlation between the state and the action. GEDT uses a graph transformer encoder to encode the GRW-CG, employing a graph attention network to aggregate information from neighboring nodes and update their representations. The encoded graph features are then fed into the sequence transformer decoder as contextual information to predict the next action in the parameter tuning process. Experimental results show that this method significantly improves the trajectory accuracy of vSLAM across multiple public datasets, demonstrating its versatility in various scenarios and simplifying the parameter tuning process.