Abstract
This work addresses a key emerging issue for building resiliency in the digital age and the advent of “smart” environments, specifically the impact of AI on human-building interactions. We present herein a concise summary of the workflows we have developed over time to simulate human-movements and dynamics of human-building interactions. We discuss the applications of human behaviour simulation in the analysis of building environments. We then discuss the approaches which use dynamic human-aware simulations in improving environment designs. We highlight the challenges in existing simulation workflows, which are usually decoupled from environment modeling tools or often tightly coupled with any specialized modeling platform. Such workflows require significant infrastructure and domain expertise, hindering the users’ abilities to seamlessly simulate, analyze, and incorporate dynamics of human-building interactions into their preferred design workflows. We introduce a democratized approach using an on-demand service-based workflow for human behaviour simulation and analysis of human-building interactions.
Introduction
Traditional Computer-Aided Design (CAD) and Building Information Modeling (BIM) tools excel at evaluating static metrics like structural integrity, energy efficiency, and lighting. However, they often fail to account for the most dynamic element of any building: the people inside it. Predicting Human-Building Interaction (HBI)—how people navigate a complex airport, evacuate during a fire, or utilize a workspace—is incredibly difficult. Existing crowd simulation tools typically require deep domain expertise and expensive hardware, and they are often "decoupled" from the actual design software architects use. This creates a high barrier to entry, meaning human-centric analytics are often ignored until late in the design process, when changes are most costly.
Methodology
The proposed framework shifts from siloed, expert-only crowd simulation tools to a "democratized," Simulation-as-a-Service (SaaS) platform. Key technical components include:
- High-Fidelity Agent-Based Modeling (ABM): Moving beyond simple "particle" models, the system uses biomechanically inspired "footstep-level" planning. This accounts for how humans actually balance and step, which significantly impacts crowd density and flow.
- Egocentric Path Planning: Agents navigate based on a simulated field of vision (egocentric fields), mimicking biological navigation rather than having "god-like" knowledge of the floor plan.
- Machine Learning (ML) for Prediction: To avoid the time-sink of running full simulations for every design tweak, the authors use Behavior Cloning and Generative Adversarial Imitation Learning (GAIL) to "learn" crowd dynamics. This allows for near-instant predictions of evacuation times and bottlenecks.
The SaaS architecture allows users to upload BIM models (like IFC files) via an API, with the server returning spatial analytics and heatmaps. This workflow is designed to be integrated directly into existing design environments.
Key Insights and Results
The research highlights that modeling "heterogeneous" walkers (people with different gaits or speeds) at the footstep level produces more accurate quantitative outcomes than treating a crowd as a uniform mass. Furthermore, in experiments with imitation learning, simple Behavior Cloning often outperformed complex Reinforcement Learning (RL) for predicting movement in dense crowds.
Key results from the application of these tools include:
- Wayfinding Efficiency: Environment layouts with fewer visual cues sometimes led to faster decision-making, suggesting that "visual noise" can hinder navigation.
- Performance Speedup: The ML-based "instant prediction" approach allows designers to see the impact of a layout change on crowd flow without rerunning expensive, minutes-long simulations.
- Automated Optimization: The tools can automatically identify optimal egress routes and suggest the best placement for signage to minimize confusion during emergencies.
Discussion and Future Work
While the framework offers a significant step toward democratization, several challenges remain. The "non-stationary problem" in multi-agent reinforcement learning—where agents' learning changes the environment for others—makes convergence difficult. Additionally, fully integrating environmental factors like temperature or sound into human movement models remains a complex frontier. Future work aims to create more "universal" crowd models that work across all building types without requiring scenario-specific training.
The core shift presented here is that the bottleneck isn't the simulation technology itself, but the access to it. Moving to a SaaS model allows architects to treat human-behavior simulation as a standard "plug-in" rather than a specialized consulting project.
Summary updated