SMART: Scalable Multi-Agent Reasoning and Trajectory Planning in Dense Environments

Huang, Heye; Yang, Yibin; Wang, Chen; Chen, Tiantian; Li, Xiaopeng; Chen, Sikai · 2025 · ArXiv.org

DOI: 10.48550/arxiv.2509.15737

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of Multi-Vehicle Trajectory Planning (MVTP) in dense environments, where the quadratic growth of pairwise collision constraints renders traditional coupled optimization methods computationally intractable. The authors introduce SMART (Scalable Multi-Agent Reasoning and Trajectory Planning), a hierarchical framework designed to decouple high-level behavior reasoning from low-level trajectory optimization. This approach aims to resolve the trade-off between exploring diverse interaction modes (homotopy classes) and ensuring real-time, kinematically feasible planning for large fleets. The SMART framework consists of two layers. The upper layer employs a centralized Priority-Based Search (PBS) to explore feasible interaction patterns. It utilizes a Synthetic Score-based Attention Network (S2AN) for warm-start initialization and Spatiotemporal Hybrid A* (STHA*) for single-vehicle planning within a joint state-time domain. This layer generates initial trajectories that resolve conflicts by branching on vehicle priority orders. The lower layer performs distributed optimization using Sequential Quadratic Programming (SQP). It constructs robust, convex safe corridors around the initial trajectories, effectively decoupling the non-convex joint problem into independent convex subproblems. By approximating vehicles with dual circles and using neighborhood search to prune irrelevant collision constraints, the method enables parallel trajectory refinement that ensures smoothness and collision avoidance. Experimental results demonstrate that SMART significantly outperforms existing baselines in scalability and speed. In simulations on 50 m × 50 m maps, SMART maintained a success rate above 90% within 1 second for up to 25 vehicles, whereas baseline methods often dropped below 50%. On larger 100 m × 100 m maps, the framework achieved over 95% success for up to 50 vehicles and remained feasible for up to 90 vehicles. Computationally, SMART was more than an order of magnitude faster than optimization-only approaches. Real-world experiments further validated the system’s practicality, achieving planning times as low as 0.014 seconds while preserving cooperative behaviors through vehicle-to-infrastructure (V2I) communication. The significance of this work lies in its ability to handle large-scale, dense multi-agent coordination in real-time. By combining learned heuristics for global behavior search with parallelizable convex optimization for local refinement, SMART overcomes the limitations of coupled methods (computational cost) and distributed methods (deadlocks). The integration of V2I support enhances safety and scalability, providing a robust solution for intelligent transportation systems requiring efficient fleet coordination in constrained spaces.

Key finding

The SMART framework achieves significantly higher success rates and faster computation times than baseline methods for multi-vehicle trajectory planning in dense environments, sustaining over 90% success within one second for up to 25 vehicles and remaining feasible for up to 90 vehicles.

Methodology

simulation_modeling

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via author_sweep_intake on 2026-05-28.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• traffic density