The Collaborative Task Sequencing and Multi-Agent Path Finding (CTS-MAPF) problem requires agents to accomplish sequences of tasks while avoiding collisions, posing significant challenges due to its combinatorial complexity. This work introduces CTS-PLL, a hierarchical framework that extends the configuration-based CTS-MAPF planning paradigm with two key enhancements: a lock agents detection and release mechanism leveraging a complete planning method for local re-planning, and an anytime refinement procedure based on Large Neighborhood Search (LNS). These additions ensure robustness in dense environments and enable continuous improvement of solution quality. Extensive evaluations across sparse and dense benchmarks demonstrate that CTS-PLL achieves higher success rates and solution quality compared with existing methods, while maintaining competitive runtime efficiency. Real-world robot experiments further demonstrate the feasibility of the approach in practice.
Figures & tables
Fig. 1: Examples of deadlock and livelock cases in CTS-PIBT. (a) Deadlock: agent a1 attempts to execute task vt1 , but agent a2 has no alternative position to yield, blocking the move. (b) Livelock: agents a1 and a2 repeatedly oscillate around each other’s goals, preventing convergence.
Fig. 2: The main process of CTS-PLL, which has four key components: joint task sequencing, path finding using Extended-PIBT, lock detection and release module and solution quality enhancement via LNS.
Scenario type
Maps
Agents N
Tasks M
Instances
Sparse
four maps
5,10,20
10,20,30,40,50
50
Dense
maze
50
60,90,120,150,180
50
TABLE I: Experimental settings for sparse and dense scenarios.
Fig. 3: Benchmark maps adopted for evaluation, including four representative scenarios: empty, random, room, and maze. These maps are used to construct sparse and dense experimental settings.
Fig. 4: Performance of CTS-PLL and baseline methods under different difficulty levels (map types, agent numbers, and task numbers). For clarity, we omit the x-axis labels in each subfigure since they share the same contents. The x-axis represents the pairs of (N,M) (agent number, task number), ranging from (5,10) to (5,50), (10,10) to (10,50), and (20,10) to (20,50). Here, CTS-PLL-v2 denotes the variant without anytime optimization, while CTS-PLL-v3 refers to the full method with anytime mechanism and LNS refinement.
Method
Success Rate ↑
# Tasks
60
90
120
150
180
CTS-PIBT
76%
68%
66%
50%
38%
CTS-PLL-v1
98%
92%
96%
92%
72%
Improvement Rate
28.95%
35.29%
45.45%
84.00%
89.47%
TABLE II: Success rate results of denser configurations on maze map.
Method
Solution Quality (Flowtime) ↓
# Tasks
60
90
120
150
180
CTS-PIBT-anytime
196.35
253.21
316.93
389.62
457.04
CTS-PLL-v3
162.00
221.89
301.85
357.99
451.34
Improvement Rate
17.49%
12.37%
4.76%
8.12%
1.25%
TABLE III: Solution quality results of denser configurations on maze map.
Fig. 5: Physical robot experiment using toio robots to validate CTS-PLL. Robots are shown as purple, cyan, and yellow squares. The corresponding colored stars indicate their goal locations, while red triangles represent task points. Black numbers above each task denote the robots assigned to it, and red numbers below show the robots that have already completed it. The top subfigure illustrates the real-world robot experiment, while the bottom subfigure shows the corresponding simulation process. The figure depicts the trajectories of the robots during the experiment.