STGN: A Spatio-Temporal Graph Network for Real-Time and Generalizable Trajectory Planning

In dynamic and unstructured environments, mobile robots need to generate safe and efficient trajectories in real time, which poses significant challenges due to the uncertainty of surrounding obstacles. To address this, this article presents a real-time obstacle avoidance trajectory planning method, built upon a spatio-temporal graph network that integrates temporal modeling with graph attention mechanisms. The proposed network captures both temporal dynamics and spatial structural dependencies in dynamic environments by integrating a temporal information module based on long short-term memory (LSTM) and a spatial module based on relational graph attention networks (RGAT). On the whole, the approach follows a two-phase pipeline. In the offline phase, a high-quality trajectory dataset is constructed to represent the heterogeneous state graph of the robot and surrounding obstacles. Then the dataset is used to train the spatio-temporal network, which learns to map environment-state graphs to optimal control commands. In the online phase, the trained network is deployed on the robot to perform real-time perception, decision-making, and control, forming a closed-loop trajectory optimization process. Extensive experiments in both simulated and real-world scenarios demonstrate that the proposed method achieves high-quality trajectory planning, robust obstacle avoidance, and fast generalization under multi-obstacle and sudden disturbance conditions, while maintaining low computational overhead. Note to Practitioners—This article addresses the generalization challenges commonly encountered in traditional supervised learning-based obstacle avoidance methods. We propose a trajectory planning framework that leverages a spatiotemporal graph network to model dynamic interactions between a robot and its surrounding obstacles. This approach enables robust and adaptable behavior in complex, changing environments by explicitly capturing both spatial and temporal dependencies. The system is implemented on a four-wheel steering (4WS) robot for experimental validation. However, its modular design ensures straightforward transferability to a wide range of mobility platforms. The proposed method requires only basic obstacle position information, readily available from standard onboard sensors such as LiDAR or radar, and does not depend on raw sensor inputs or semantic maps, making it highly suitable for real-world deployment. It can be easily integrated into existing perception–planning–control pipelines, and future extensions may focus on incorporating richer semantic information and expanding to more diverse obstacle types.