Editable Map-Conditioned Trajectory Generation for Human Mobility Simulation

Takayuki Mizuno; Shouji Fujimoto; Mikito Hiruki; Atushi Ishikawa

cs.LGSep 26, 2026

Editable Map-Conditioned Trajectory Generation for Human Mobility Simulation

Authors: Takayuki Mizuno, Shouji Fujimoto, Mikito Hiruki, Atushi Ishikawa

Organizations: National Institute of Informatics Tokyo, Japan · Kanazawa Gakuin University Kanazawa, Ishikawa, Japan · The University of Tokyo Tokyo, Japan

Abstract

Geospatial simulation of infrastructure interventions requires mobility generators that respond directly to edited maps, yet many data-driven generators do not expose the map as an editable condition. We formulate this task as map-conditioned autoregressive generation of human mobility: a road raster conditions a decoder that emits nominal 31.25 m mesh-cell tokens at one-minute intervals. The mesh-local vocabulary supports held-out and locally edited maps without retraining or vocabulary changes. We instantiate a ResNet-50 visual-prefix configuration and a Vision Transformer (ViT) cross-attention configuration, trained from scratch on 87,400 smartphone-derived trajectories from 874 meshes in Ishikawa Prefecture, Japan; 219 meshes are held out. We evaluate map sensitivity by comparing correct-map and within-split shuffled-map generations with held-out real trajectories. On the 110-mesh test split, for the ResNet-50 configuration, correct-map generations are closer than shuffled-map generations on 60% of meshes under Hausdorff-based energy distance (p = 0.021), while DTW is directional but inconclusive (57%, p = 0.074); correlation with real density is 0.38 with the correct map versus 0.01 with shuffled maps. The ViT configuration shows weaker trajectory-level sensitivity and smaller density gains. An illustrative bridge-removal edit changes generated continuations without retraining. Together, these results support the feasibility of editable-map human-mobility simulation.

Figures & tables

Figure 1. Overview of the input and output. Top: standard image captioning maps an image to word tokens. Bottom: the proposed model maps a road raster $I_{m}$ to mesh-cell tokens. The red trajectory is not part of the input; it is overlaid only after the generated tokens are converted to geographic cell centroids. All parameters are learned from scratch. A schematic compares standard image captioning with the proposed map-conditioned trajectory generator. A road raster is encoded and decoded into mesh-cell tokens, which are converted to geographic centroids and plotted as a trajectory.

Tier	Category	Width (m)	Weight $w_{h}$
Arterial (L3)	motorway	20	0.45
	motorway_link	20	0.16
	trunk	20	0.41
	trunk_link	20	0.19
	primary	20	0.37
	primary_link	20	0.16

Table 1. OpenStreetMap highway categories with their rendered road widths and brightness weights

w_{h}

. Weights are estimated by

L_{2}

regression against the training mobility-density images of Equation 4 ; categories are grouped into arterial (L3), major local (L2), residential (L1), and auxiliary tiers.

Panel of Figure 2 — Figure 2. Training and validation loss. The selected checkpoints are epoch $4$ for the CNN configuration and epoch $5$ for the ViT configuration. Two plots show training and validation loss over eight epochs for CNN- and ViT-conditioned trajectory generators.

Panel of Figure 3 — Figure 3. Map-only comparison of $100$ trajectories per configuration. From left: conditioning map, CNN generations, and ViT generations. (a) Synthetic circular road. (b) OSM-derived urban road image. Two rows compare a conditioning map with one hundred CNN- and ViT-generated trajectories on a circular road and an urban road raster.

		CNN		ViT
Distance	Split	$\hat{p}$	$p$ -value	$\hat{p}$	$p$ -value
Hausdorff	Train	$0.71$	$10^{-38}$	$0.47$	$0.944$
	Validation	$0.62$	$0.0079$	$0.44$	$0.895$
	Test	$0.60$	$0.021$	$0.50$	$0.538$
DTW	Train	$0.62$	$10^{-12}$	$0.49$	$0.740$
	Validation	$0.58$	$0.051$	$0.47$	$0.750$

Table 2. Trajectory-level map sensitivity. Positive ratio

\hat{p}

and one-sided binomial

p

-value for

\Pr(\Delta D_{E,m}>0)>1/2

. Values above

0.5

favor the correct map on a majority of meshes. Because the

p

-value is a deterministic function of the number of positive meshes, the identical DTW test entries for CNN and ViT (

\hat{p}=0.57

p=0.074

) reflect the same count of positive meshes and are not a transcription error.

Panel of Figure 4 — Figure 4. Per-mesh improvement $\Delta D_{E,m}$ on $219$ held-out meshes. Rows: Hausdorff and DTW; columns: CNN and ViT. Positive values favor the correct map over the shuffled map. Four histograms show correct-map improvement for CNN and ViT models under Hausdorff and DTW trajectory distances.

	CNN				ViT
Split	$\overline{\mathrm{corr}}^{+}$	$\overline{\mathrm{corr}}^{-}$	$\hat{p}$	$p$ -value	$\overline{\mathrm{corr}}^{+}$	$\overline{\mathrm{corr}}^{-}$	$\hat{p}$	$p$ -value
Train	0.41	0.00	$0.97$	$10^{-214}$	0.15	0.00	$0.79$	$10^{-69}$
Validation	0.41	0.03	$0.96$	$10^{-26}$	0.13	0.01	$0.77$	$10^{-9}$
Test	0.38	0.01	$0.94$	$10^{-13}$	0.14	0.00	$0.75$	$10^{-8}$

Table 3. Density-level map sensitivity. Mean heatmap correlations for correct (

\overline{\mathrm{corr}}^{+}

) and shuffled (

\overline{\mathrm{corr}}^{-}

) maps, positive ratio

\hat{p}

, and one-sided binomial

p

-value.

Figure 5. Illustrative edited-map intervention. CNN-generated continuations (red; $50$ from each of two prefixes) under the original map (left) and a map with the northwestern bridge removed (right). Only the input raster changes. Both panels show a single $1\,\mathrm{km}$ mesh; generated positions are the centroids of its $32\times 32$ mesh-local cells (nominal $31.25\,\mathrm{m}$ ), which is why the paths appear much finer than the mesh itself. Side-by-side maps show generated trajectories before and after one bridge is removed from the road raster.

Trajectory Generation

Trajectory generation creates plausible, goal-directed paths and motion sequences for people, vehicles, and robots.

See all 181 papers

ActivityEditor: Learning to Synthesize Physically Valid Human Mobility
Apr 7, 2026Chenjie Yang, Yutian Jiang, Anqi Liang +3Trajectory Generation LLM Agent Trajectory Synthesis
Southwest Jiaotong University · HKUST (Guangzhou) · Shanghai Jiao Tong University +2
TrajGenAgent: A Hierarchical LLM Agent for Human Mobility Trajectory Generation
Jun 10, 2026Siyu Li, Toan Tran, Lingyi Zhao +2Trajectory Generation Human Behavior Simulation
Dept. of Computer Science, Emory University, Atlanta, GA, USA · Novateur Research Solutions, Ashburn, VA, USA
TrajDLM: Topology-Aware Block Diffusion Language Model for Trajectory Generation
May 11, 2026Wilson Wongso, Lihuan Li, Arian Prabowo +4Trajectory Generation Synthetic Data Generation
University of New South Wales · Hong Kong University of Science and Technology (Guangzhou)

Apr 7, 2026J/K move · Enter open · S save

ActivityEditor: Learning to Synthesize Physically Valid Human Mobility

Chenjie Yang, Yutian Jiang, Anqi Liang, Wei Qi+2

Southwest Jiaotong University · HKUST (Guangzhou) · Shanghai Jiao Tong University+2

Trajectory Generation LLM Agent Trajectory Synthesis

Open paper AskPDF arXiv

Human mobility modeling is indispensable for diverse urban applications. However, existing data-driven methods often suffer from data scarcity, limiting their applicability in regions where historical trajectories are unavailable or restricted. To bridge this gap, we propose ActivityEditor, a novel dual-LLM-agent framework designed for zero-shot cross-regional trajectory generation. Our framework decomposes the complex synthesis task into two collaborative stages. Specifically, an intention-based agent, which leverages demographic-driven priors to generate structured human intentions and coarse activity chains to ensure high-level socio-semantic coherence. These outputs are then refined by editor agent to obtain mobility trajectories through iteratively revisions that enforces human mobility law. This capability is acquired through reinforcement learning with multiple rewards grounded in real-world physical constraints, allowing the agent to internalize mobility regularities and ensure high-fidelity trajectory generation. Extensive experiments demonstrate that ActivityEditor achieves superior zero-shot performance when transferred across diverse urban contexts. It maintains high statistical fidelity and physical validity, providing a robust and highly generalizable solution for mobility simulation in data-scarce scenarios. Our code is available at: https://anonymous.4open.science/r/ActivityEditor-066B.

Abstract

Trajectory Generation

Explore similar work

ActivityEditor: Learning to Synthesize Physically Valid Human Mobility

TrajGenAgent: A Hierarchical LLM Agent for Human Mobility Trajectory Generation

TrajDLM: Topology-Aware Block Diffusion Language Model for Trajectory Generation