Language-Grounded Multi-Agent Planning for Personalized and Fair Participatory Urban Sensing

Xusen Guo; Mingxing Peng; Hongliang Lu; Hai Yang; Jun Ma; Yuxuan Liang

cs.AIMar 25, 2026

Language-Grounded Multi-Agent Planning for Personalized and Fair Participatory Urban Sensing

Authors: Xusen Guo, Mingxing Peng, Hongliang Lu, Hai Yang, Jun Ma, Yuxuan Liang

Organizations: The Hong Kong University of Science and Technology (Guangzhou) · The Hong Kong University of Science and Technology

Abstract

Participatory urban sensing leverages human mobility for large-scale urban data collection, yet existing methods typically rely on centralized optimization and assume homogeneous participants, resulting in rigid assignments that overlook personal preferences and heterogeneous urban contexts. We propose MAPUS, an LLM-based multi-agent framework for personalized and fair participatory urban sensing. In our framework, participants are modeled as autonomous agents with individual profiles and schedules, while a coordinator agent performs fairness-aware selection and refines sensing routes through language-based negotiation. Experiments on real-world datasets show that MAPUS achieves competitive sensing coverage while substantially improving participant satisfaction and fairness, promoting more human-centric and sustainable urban sensing systems.

Figures & tables

Figure 1: Comparison of paradigms for PUS problem. (a) Conventional centralized PUS relies on simplified assumptions about participants and sensing regions. (b) MAPUS models participants as autonomous agents with personalized preferences while incorporating heterogeneous urban context. (c) AgentSense ( Guo et al., 2025b ) uses LLM agents as workflow components to iteratively refine a baseline PUS solution under disturbances. (d) MAPUS adopts a coordinator-and-participants multi-agent paradigm, enabling decentralized, personalized, and fairness-aware planning.

Figure 2: Overview of proposed MAPUS. It operates in three stages: (a) task dispatch and preference-aware route generation, (b) fairness-aware participant selection, and (c) negotiation-based route refinement.

Dataset	Scale	# Regions	$N$	$B$	$H$ (min)
T-Drive	Small	64 (8 $\times 8$ )	20	40	120
	Medium	256 (16 $\times 16$ )	40	60	240
	Large	1024 (32 $\times 32$ )	60	100	360
Grab-Posisi	Small	32 (8 $\times$ 4)	15	40	40
	Medium	128 (16 $\times 8$ )	30	60	80
	Large	512 (32 $\times 16$ )	45	100	160

Table 1: Experimental configurations under different scales for the two datasets.

\#\text{Regions}

denotes the number of spatial sensing regions,

N

the number of candidate participants,

B

the total budget, and

H

the temporal span horizon of the sensing task.

Method	Task Scales
	Small		Medium		Large
	Coverage	PSS	Coverage	PSS	Coverage	PSS
RN	4.557 $\pm$ 8.38e-3	0.632 $\pm$ 9.60e-5	4.626 $\pm$ 1.46e-2	0.605 $\pm$ 4.65e-5	4.974 $\pm$ 3.68e-2	0.594 $\pm$ 3.10e-4
TVPG	4.563 $\pm$ 1.34e-2	0.634 $\pm$ 9.68e-5	4.753 $\pm$ 2.27e-2	0.610 $\pm$ 6.62e-5	5.067 $\pm$ 1.34e-2	0.590 $\pm$ 2.14e-4
TCPG	4.594 $\pm$ 7.31e-3	0.631 $\pm$ 4.58e-5	5.084 $\pm$ 1.13e-2	0.621 $\pm$ 3.48e-5	5.043 $\pm$ 1.71e-2	0.602 $\pm$ 1.17e-4
MSA	4.662 $\pm$ 5.60e-3	0.630 $\pm$ 3.60e-5	5.065 $\pm$ 9.24e-3	0.622 $\pm$ 8.63e-5	5.131 $\pm$ 1.70e-2	0.598 $\pm$ 1.80e-4

Table 2: Overall performance comparison on the T-Drive dataset (20 task instances per scale). Values denote averages, with variances shown in small numbers. The largest variance per column is highlighted in red.

Method	Task Scales
	Small		Medium		Large
	Coverage	PSS	Coverage	PSS	Coverage	PSS
RN	4.236 $\pm$ 4.23e-3	0.727 $\pm$ 4.29e-5	4.710 $\pm$ 1.18e-2	0.696 $\pm$ 6.30e-5	5.399 $\pm$ 1.11e-2	0.648 $\pm$ 5.13e-5
TVPG	4.342 $\pm$ 4.10e-3	0.714 $\pm$ 4.59e-5	4.737 $\pm$ 1.32e-2	0.695 $\pm$ 2.04e-4	5.406 $\pm$ 2.62e-2	0.640 $\pm$ 5.74e-5
TCPG	4.232 $\pm$ 2.83e-3	0.730 $\pm$ 2.86e-5	4.806 $\pm$ 1.14e-2	0.703 $\pm$ 6.40e-5	5.587 $\pm$ 8.39e-3	0.658 $\pm$ 3.56e-5
MSA	4.372 $\pm$ 2.56e-3	0.744 $\pm$ 3.54e-5	4.900 $\pm$ 4.15e-3	0.699 $\pm$ 9.28e-5	5.586 $\pm$ 8.20e-3	0.655 $\pm$ 4.07e-5

Table 3: Overall performance comparison on the Grab-Posisi dataset (20 task instances per scale). Values denote averages, with variances shown in small numbers. The largest variance per column is highlighted in red.

Figure 3: Ablation study on the T-Drive dataset. We compare four variants of MAPUS across three different settings: w/o PRG, w/o FPS, w/o NRR, and the full model. Top: coverage utility. Bottom: path satisfaction score (PSS).

Metric	T-Drive		Grab-Posisi
Metric	Before Nego.	After Nego.	Before Nego.	After Nego.
Avg. Overlap (%)	3.482	1.267	5.418	3.843
Avg. Coverage	4.977	5.115	4.770	4.890
Avg. Entropy	2.377	2.477	2.498	2.584
Avg. Count	149.3	149.3	132.1	132.1
Avg. PSS	0.880	0.875	0.838	0.834

Table 4: Effect of negotiation-based route refinement under the Medium setting (20 tasks). We report overlap, coverage utility, entropy, covered count, and PSS before and after negotiation on T-Drive and Grab-Posisi.

Figure 4: Fairness-aware participant selection analysis on the T-Drive dataset under the Medium setting. (a) Per-worker selection counts over 60 tasks, (b) CDF of selection counts, (c) data coverage utility, and (d) path satisfaction score.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 5 — Figure 5: Preference-aware route generation prompt.

Panel of Figure 6 — Figure 6: Prompt template used for LLM-based tie-breaking in fairness-aware participant selection..

Panel of Figure 7 — Figure 7: Prompt for the coordinator agent in the negotiation process.

Panel of Figure 8 — Figure 8: Prompt for the participant agent in the negotiation process.

Figure 9: Visualization of land-use and crime distribution in task regions. (a) Land-use distribution in the T-Drive task area derived from aggregated POI categories. (b) Spatial distribution of crime counts in the T-Drive task area aggregated over grid cells. (c) Land-use distribution in the Grab-Posisi task area.

Panel of Figure 10 — Figure 10: Template used to construct synthetic participant profiles.

Figure 11: Performance trade-offs of different base LLMs under the Medium setting (20 tasks). The x-axis shows path satisfaction score (PSS) and the y-axis shows coverage utility. Bubble size indicates runtime. Results are shown for T-Drive (left) and Grab-Posisi (right).

Method	Avg. runtime on T-Drive (s)			Avg. runtime on Grab-Posisi (s)
Method	Small	Medium	Large	Small	Medium	Large
RN	6.754e-4	4.453e-4	1.142e-3	2.999e-4	9.423e-4	1.568e-3
TVPG	0.010	0.048	0.472	0.0039	0.0212	0.266
TCPG	0.012	0.112	1.155	0.0044	0.0394	0.388
MSA	0.040	0.077	0.217	0.0319	0.0616	0.157
MSAGI	0.089	0.322	2.442	0.0509	0.168	1.502

Appendix

Table 5: Runtime comparison across datasets and task scales. The main values denote the average runtime (seconds), and the lower-right numbers indicate the variance across 20 task instances.

Figure 12: Case studies of route adaptation under dynamic disturbances. (a) A traffic accident causing grid [5, 2] blockage in the T-Drive dataset. (b) A high-priority data collection task in grid [14, 4] from the Grab-Posisi dataset. In each case, the coordinator agent (CA) broadcasts dynamic information to participant agents (PAs). PAs evaluate whether their routes are affected and adjust accordingly if necessary. We report the coverage utility and path satisfaction score for the initial and refined solutions, with the areas modified in response to the dynamic disturbances highlighted by red arrows.

Explore similar work

Sentinel: Embodied Cooperative Spatial Reasoning and Planning
May 25, 2026Xiangye Lin, Hongxin Zhang, Ruxi Deng +2AI Agent Benchmarks Multi-Agent Planning
University of Massachusetts Amherst · University of Illinois Urbana-Champaign
UrbanHuRo: A Two-Layer Human-Robot Collaboration Framework for the Joint Optimization of Heterogeneous Urban Services
Mar 4, 2026Tonmoy Dey, Lin Jiang, Zheng Dong +1Human-Robot Collaboration RL for Robotics
Department of Computer Science, Florida State University. · Department of Computer Science, Wayne State University.
Multi-Agent Collaborative Reasoning with Tool-Augmented Evidence for Urban Region Profiling
Jul 15, 2026Xixuan Hao, Yutian Jiang, Jiabo Liu +4Multi-Agent Reasoning Multimodal Agents
The Hong Kong University of Science and Technology (Guangzhou) Guangzhou, China · University of Washington Seattle, USA · Chang’an University Xi’an, China +1

May 25, 2026J/K move · Enter open · S save

Sentinel: Embodied Cooperative Spatial Reasoning and Planning

Xiangye Lin, Hongxin Zhang, Ruxi Deng, Qinhong Zhou+1

University of Massachusetts Amherst · University of Illinois Urbana-Champaign

AI Agent Benchmarks Multi-Agent Planning Multi-Agent Communication Multi-Agent Path Finding Spatial Reasoning

Open paper AskPDF arXiv

In this work, we study Cooperative Spatial Intelligence, the ability of decentralized embodied agents to coordinate effectively under dynamic environmental constraints across city-scale outdoor domains. We introduce Sentinel Challenge, a benchmark where multiple decentralized embodied agents must communicate in natural language to agree on a mutually safe and convenient meeting point within large, city-scale outdoor environments. Each agent must then navigate safely while avoiding dynamic sentinels patrolling the area, using a tool that provides coarse spatial information. To address this, we propose CoSaR (Cooperative Spatial Reasoning and Planning), a framework that bridges the high-level communication and planning abilities of foundation models with the precision of classical spatial navigation algorithms. CoSaR enables agents to exchange situational updates, reason over evolving spatial constraints, and collaboratively replan trajectories. Evaluated across 14 city-level scenes with 3-5 agents, CoSaR consistently leads to faster gathering, shorter path lengths, and improved safety. Our results demonstrate that integrating dynamic communication with spatial reasoning is essential for robust multi-agent cooperation. By formalizing this new setting and providing a scalable benchmark, we aim to build a foundation for advancing cooperative spatial intelligence in embodied multi-agent systems. Code and challenge are available at https://github.com/UMass-Embodied-AGI/Sentinel.

Abstract

Explore similar work

Sentinel: Embodied Cooperative Spatial Reasoning and Planning

UrbanHuRo: A Two-Layer Human-Robot Collaboration Framework for the Joint Optimization of Heterogeneous Urban Services

Multi-Agent Collaborative Reasoning with Tool-Augmented Evidence for Urban Region Profiling