Scholarlink

<p>Modern warehouse automation increasingly relies on multi-agent path coordination under heterogeneous, dynamic constraints. Most existing MAPF approaches treat all obstacles uniformly, overlooking the fundamental distinction between communicable agent peers and non-communicable moving obstacles (e.g., forklifts, hand-carts). This paper presents a heterogeneous dynamic obstacle-aware multi-AGV planning framework comprising three integrated components: (1) a Transformer-based trajectory predictor for external obstacles with calibrated uncertainty quantification (87% action accuracy, ADE=1.14); (2) an intent-aware spatiotemporal risk map that differentiates threat levels based on obstacle type and predicted intent; (3) a dual-layer planning architecture integrating risk-aware A* for collision avoidance with an iterative conflict resolution mechanism for multi-agent coordination. Experiments on 13 benchmark maps demonstrate the approach outperforms baselines (PBS, LNS) across 5–10 agent scenarios in success rate, collision count, and makespan. The method achieves 25–370<InlineEquation ID="IEq1"><EquationSource Format="TEX">\(\times \)</EquationSource><EquationSource Format="MATHML"><math><mo>×</mo></math></EquationSource></InlineEquation> computational speedup over prediction-enhanced baselines while maintaining superior path quality. Ablation studies validate each module’s contribution, demonstrating that coordination mechanisms dominate performance in high-density scenarios.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Intent-aware spatiotemporal risk field for multi-AGV path planning with heterogeneous dynamic obstacles

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月2日

查看全文
  • Yunkai Qiu,
  • Yuanhong Liu

摘要

Modern warehouse automation increasingly relies on multi-agent path coordination under heterogeneous, dynamic constraints. Most existing MAPF approaches treat all obstacles uniformly, overlooking the fundamental distinction between communicable agent peers and non-communicable moving obstacles (e.g., forklifts, hand-carts). This paper presents a heterogeneous dynamic obstacle-aware multi-AGV planning framework comprising three integrated components: (1) a Transformer-based trajectory predictor for external obstacles with calibrated uncertainty quantification (87% action accuracy, ADE=1.14); (2) an intent-aware spatiotemporal risk map that differentiates threat levels based on obstacle type and predicted intent; (3) a dual-layer planning architecture integrating risk-aware A* for collision avoidance with an iterative conflict resolution mechanism for multi-agent coordination. Experiments on 13 benchmark maps demonstrate the approach outperforms baselines (PBS, LNS) across 5–10 agent scenarios in success rate, collision count, and makespan. The method achieves 25–370\(\times \)× computational speedup over prediction-enhanced baselines while maintaining superior path quality. Ablation studies validate each module’s contribution, demonstrating that coordination mechanisms dominate performance in high-density scenarios.