Hybrid FSM-MARL Crowd Simulation for Retail Visualization in Unity

The ability to simulate realistic crowd movement is a cornerstone requirement across diverse fields, including urban planning, emergency preparedness, virtual environment design, and visualizing retail spaces. However, traditional simulation techniques frequently prove inadequate when it comes to representing the subtle and varied actions of individuals. This inadequacy is visible in settings such as retail environments, where shopper decisions are shaped by a mix of various factors. Our paper introduces a novel hybrid framework designed to simulate these complex behaviours. It achieves this by merging the structured control of Finite State Machines (FSM) with the adaptive learning power of Multi-Agent Reinforcement Learning (MARL). The goal is to generate truly lifelike and diverse shopper behaviours specifically within a Unity-based retail simulation environment. Within our system, FSMs dictate the primary stages of an agent’s path – the sequence of Entering, Shopping, Checking Out, and Exiting. Yet, the more subtle decision-making during the actual shopping phase is handled by MARL, implemented using Proximal Policy Optimization (PPO) via Unity ML-Agents. To model the diversity seen in real shoppers, we developed and trained three distinct behavioural archetypes: the focused Goal-Oriented shopper, the easily swayed Impulsive shopper, and the freely browsing Explorative (or ‘Wanderer’) shopper. The generated data shows that these shopper types made distinct decisions, resulting in unique behaviours, demonstrating that the system can produce realistic and varied shopper behaviour.