Selected work

Representative Papers

TAPDreamer: Transferable Adversarial Patches for World Action Models

Oct 05, 2026

This study addresses the vulnerability of visual encoders in world models to localized adversarial attacks and the reliance of existing methods on target outputs. We propose a universal adversarial patch generation method that generalizes across tasks and architectures. By leveraging publicly available encoders, our approach constructs fixed local perturbations through maximizing global representation shifts. It reveals how attention–value interactions enable stable shift broadcasting, facilitating black-box attacks without querying the target policy. Efficient generation is achieved by optimizing the global L1 distance using only six frames from the source task. Evaluated on the LIBERO and RoboTwin benchmarks, the proposed patch reduces task success rates to 0%, underscoring the urgent need to secure shared visual encoders.

0 citationsRead paper

Proxemics and Permeability of the Pedestrian Group

Oct 30, 2025

This study investigates how interpersonal interactions among pedestrian groups shape collective proxemic structures—specifically, spatial zoning around groups, regional occupancy frequency, and “clearing behavior” exhibited by individuals approaching groups. Method: Leveraging large-scale, naturalistic video observation data, we integrate spatial point pattern analysis with behavioral sequence statistics to construct a dynamic, quantifiable model of group-peripheral permeability. Contribution/Results: We identify three concentric, progressively restrictive social zones surrounding groups—beyond the public zone—revealing that individual approach behaviors are infrequent, brief, and occur in incremental, stepwise penetration. This provides the first empirical evidence for a stratified proxemic structure in group contexts. The findings confirm systematic regulatory effects of social norms on collective spatial organization and advance theoretical understanding of the spatial logic underlying group movement. These insights offer novel foundations for agent-based crowd simulation and evidence-informed design of public spaces.

0 citationsRead paper

Life-cycle Modeling and the Walking Behavior of the Pedestrian-Group as an Emergent Agent: With Empirical Data on the Cohesion of the Group Formation

Oct 30, 2025

This study investigates the life-cycle evolution of pedestrian groups as emergent intelligent agents and its influence on walking behavior, specifically addressing the relationship between dynamic group cohesion and collective intention formation. Method: We propose a state-transition-based group life-cycle model and develop a data-driven group morphology–behavior coupling model by integrating trajectory extraction from surveillance videos, expert-annotated group relationships and events, and quantitative clustering analysis. Contribution/Results: We systematically identify, for the first time, empirical correlations among cohesion dynamics, enhanced agentivity, and morphological state transitions; further, we derive generalizable abstract walking pattern sequences characterizing group locomotion. These findings provide computationally tractable, interpretable, and structurally grounded behavioral primitives—enabling high-fidelity group simulation modeling and principled design of human–machine collaborative interaction systems.

0 citationsRead paper

Recent publications

Latest Papers

TAPDreamer: Transferable Adversarial Patches for World Action Models

Oct 05, 2026

This study addresses the vulnerability of visual encoders in world models to localized adversarial attacks and the reliance of existing methods on target outputs. We propose a universal adversarial patch generation method that generalizes across tasks and architectures. By leveraging publicly available encoders, our approach constructs fixed local perturbations through maximizing global representation shifts. It reveals how attention–value interactions enable stable shift broadcasting, facilitating black-box attacks without querying the target policy. Efficient generation is achieved by optimizing the global L1 distance using only six frames from the source task. Evaluated on the LIBERO and RoboTwin benchmarks, the proposed patch reduces task success rates to 0%, underscoring the urgent need to secure shared visual encoders.

0 citationsRead paper

Proxemics and Permeability of the Pedestrian Group

Oct 30, 2025

This study investigates how interpersonal interactions among pedestrian groups shape collective proxemic structures—specifically, spatial zoning around groups, regional occupancy frequency, and “clearing behavior” exhibited by individuals approaching groups. Method: Leveraging large-scale, naturalistic video observation data, we integrate spatial point pattern analysis with behavioral sequence statistics to construct a dynamic, quantifiable model of group-peripheral permeability. Contribution/Results: We identify three concentric, progressively restrictive social zones surrounding groups—beyond the public zone—revealing that individual approach behaviors are infrequent, brief, and occur in incremental, stepwise penetration. This provides the first empirical evidence for a stratified proxemic structure in group contexts. The findings confirm systematic regulatory effects of social norms on collective spatial organization and advance theoretical understanding of the spatial logic underlying group movement. These insights offer novel foundations for agent-based crowd simulation and evidence-informed design of public spaces.

0 citationsRead paper

Life-cycle Modeling and the Walking Behavior of the Pedestrian-Group as an Emergent Agent: With Empirical Data on the Cohesion of the Group Formation

Oct 30, 2025

This study investigates the life-cycle evolution of pedestrian groups as emergent intelligent agents and its influence on walking behavior, specifically addressing the relationship between dynamic group cohesion and collective intention formation. Method: We propose a state-transition-based group life-cycle model and develop a data-driven group morphology–behavior coupling model by integrating trajectory extraction from surveillance videos, expert-annotated group relationships and events, and quantitative clustering analysis. Contribution/Results: We systematically identify, for the first time, empirical correlations among cohesion dynamics, enhanced agentivity, and morphological state transitions; further, we derive generalizable abstract walking pattern sequences characterizing group locomotion. These findings provide computationally tractable, interpretable, and structurally grounded behavioral primitives—enabling high-fidelity group simulation modeling and principled design of human–machine collaborative interaction systems.

0 citationsRead paper