ST-DiffPlanner: A Safety-Enhanced Topology-Aware Diffusion Planner for Global Path Planning
Jiaquan Yan, Fang Zhao, Huiyu Yuan, Yushi Chen, Long Wang, Dan Luo, Haiyong Luo
AI summary
Problem
Existing diffusion-based global path planners lack explicit safety constraints, often generating obstacle-penetrating trajectories in complex maze-like environments, while traditional and learning-enhanced methods suffer from unstable real-time performance and poor cross-scenario generalization.
Approach
The method integrates a Graph Neural Network (GNN) to extract topological features that guide a diffusion model, combined with a topology anchor-based safety loss for training and a bidirectional safety projection strategy during inference to ensure collision-free, continuous trajectories.
Key results
- GNN-integrated topology-aware diffusion framework for global spatial understanding
- Topology anchor-based differentiable safety loss for stable training and safety boundary recognition
- Topology-aware bidirectional safety projection strategy ensuring trajectory safety during inference
- 96.9% average trajectory generation success rate with strong cross-scenario and cross-platform generalization
Why it matters
It provides a robust, real-time global path planning solution for robots operating in complex, narrow, and maze-like environments where safety and generalization are critical.
Abstract
In complex environments, traditional path plan- ning methods rely on manually defined models, requiring tedious adjustments under varying scenarios or constraints. They also suffer from unstable time overhead and exponen- tially increasing computational costs as environmental com- plexity grows. Deep learning-enhanced methods, while op- timizing decisions via neural networks, remain constrained by explicit search/sampling frameworks—this leads to unsta- ble real-time performance and failure to capture real-world trajectory distributions. In contrast, diffusion-based planning directly learns trajectory distributions from data, offering predictable inference latency via fixed inversion steps and inherent support for multimodal solutions. However, its lack of explicit safety constraints often leads to trajectory safety issues, resulting in planning failures. To address these lim- itations, this paper proposes ST-DiffPlanner, a global path planner following the pipeline of “topology cognition—direction focusing—trajectory generation”. It introduces three targeted optimizations: (1) leveraging topological awareness to con- strain the diffusion model to focus on collision-free regions; (2) optimizing inference-phase projection to ensure trajectory continuity and safe distances from obstacles; (3) designing a topology anchor-based safety loss to enhance model safety and training stability. Experimental results demonstrate that ST-DiffPlanner exhibits strong generalization across multiple scenarios and modalities, accurately capturing environmental features and learning task-compliant trajectory characteristics. Our method achieves an average trajectory generation success rate of 96.9%, significantly outperforming baseline methods. Moreover, validation in both simulated and real-world robot platforms confirms its applicability across different systems.