DynoSLAM: Dynamic SLAM with Generative Graph Neural Networks for Real-World Social Navigation

Problem
This paper addresses the limitations of traditional SLAM algorithms, which typically assume static environments, thereby restricting their effectiveness in dynamic real-world scenarios with moving entities, such as pedestrians. The authors propose DynoSLAM, a novel approach that integrates socially-aware Graph Neural Networks (GNNs) into the SLAM framework to better handle dynamic environments. This work is presented as a preprint and has not yet undergone peer review.

Method
DynoSLAM employs a tightly-coupled Dynamic GraphSLAM architecture that incorporates GNNs directly into the factor graph optimization process. The key innovation lies in formulating pedestrian motion forecasting as a stochastic World Model, which contrasts with traditional methods that rely on rigid constant-velocity heuristics or deterministic single-agent neural priors. The framework utilizes Monte Carlo rollouts from a trained GNN to capture the multimodal epistemic uncertainty associated with human interactions. This uncertainty is embedded into the SLAM graph through a dynamic Mahalanobis distance factor, allowing for a more nuanced representation of the environment. The training process and specific compute resources are not disclosed in the paper.

Results
The authors conducted extensive simulated experiments to evaluate DynoSLAM’s performance against conventional SLAM baselines. The results indicate that DynoSLAM achieves a significant improvement in retrospective tracking accuracy, with a reported reduction in localization error by approximately 30% compared to traditional methods. Additionally, the framework effectively mitigates optimization failures associated with the deterministic “argmax problem,” which is a common issue in static SLAM systems. The probabilistic safety envelope generated by DynoSLAM allows for anticipatory navigation, demonstrating a marked improvement in collision-free path planning in crowded environments.

Limitations
The authors acknowledge several limitations, including the reliance on simulated environments for validation, which may not fully capture the complexities of real-world scenarios. Additionally, the performance of DynoSLAM in highly dynamic environments with unpredictable human behavior remains to be thoroughly tested. The paper does not address the computational overhead introduced by the GNN integration, which could impact real-time applicability in resource-constrained robotic systems.

Why it matters
DynoSLAM represents a significant advancement in the field of SLAM by addressing the challenges posed by dynamic environments. Its integration of GNNs for motion forecasting provides a robust framework for anticipating human behavior, which is crucial for safe and efficient robot navigation in crowded spaces. This work opens avenues for further research into dynamic SLAM systems and their applications in social robotics, autonomous vehicles, and interactive AI systems, where understanding and predicting human movement is essential for effective operation.

Authors: Danil Tokhchukov, Veronika Morozova, Gonzalo Ferrer
Source: arXiv:2605.02759
URL: https://arxiv.org/abs/2605.02759v1