Rapid and Safe Trajectory Planning over Diverse Scenes through Diffusion Composition

2025-07-06

AI Summary

To address the inherent trade-off between safety and real-time performance in trajectory planning for complex environments, this paper proposes a fast and safe trajectory planning framework based on state-space diffusion models. Methodologically, it employs low-dimensional vehicle state modeling to reduce computational overhead; designs a diffusion model ensemble mechanism to achieve robust safety generalization across diverse—包括 unseen—scenarios; and integrates a lightweight, rule-driven safety filter to select optimal collision-free trajectories while ensuring control feasibility. Experiments on the F1TENTH physical platform demonstrate millisecond-level inference latency, alongside significant improvements in trajectory safety and stability, thereby overcoming the efficiency–safety trade-off limitation of conventional approaches. Key contributions include: (1) diffusion modeling directly in the vehicle state space, (2) an ensemble-based generalization mechanism for cross-scenario safety transfer, and (3) a hybrid rule-learning paradigm for safety-critical trajectory filtering.

Technology Category

• Planning, Routing, and Scheduling: Replanning and Plan Repair
• Humans and AI: Human-Aware Planning and Behavior Prediction
• Computer Vision: Diffusion Models for Vision

Application Category

• Security and Privacy: Large-scale security measurements
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Safe trajectory planning remains a significant challenge in complex environments, where traditional methods often trade off computational efficiency for safety. Comprehensive obstacle modeling improves safety but is computationally expensive, while approximate methods are more efficient but may compromise safety. To address this issue, this paper introduces a rapid and safe trajectory planning framework based on state-based diffusion models. Leveraging only low-dimensional vehicle states, the diffusion models achieve notable inference efficiency while ensuring sufficient collision-free characteristics. By composing diffusion models, the proposed framework can safely generalize across diverse scenarios, planning collision-free trajectories even in unseen scenes. To further ensure the safety of the generated trajectories, an efficient, rule-based safety filter is proposed, which selects optimal trajectories that satisfy both sufficient safety and control feasibility from among candidate trajectories. Both in seen and unseen scenarios, the proposed method achieves efficient inference time while maintaining high safety and stability. Evaluations on the F1TENTH vehicle further demonstrate that the proposed method is practical in real-world applications. The project page is at: https://rstp-comp-diffuser.github.io/.

Problem

• Safe trajectory planning in complex environments
• Balancing computational efficiency and safety
• Generalizing across diverse and unseen scenarios

Innovation

• Diffusion models for efficient trajectory planning
• Composition of diffusion models across scenarios
• Rule-based safety filter for optimal trajectories