Scenario Generalization of Data-driven Imitation Models in Crowd Simulation

Gang Qiao, Honglu Zhou, Mubbasir Kapadia, Sejong Yoon, Vladimir Pavlovic

Abstract

Crowd simulation, the study of the movement of multiple agents in complex environments, presents a unique application domain for machine learning. One challenge in crowd simulation is to imitate the movement of expert agents in highly dense crowds. This work investigates how training data and training methods affect imitators when these models are applied to novel scenarios. A general imitation model is represented by applying either Behavior Cloning (BC) or Generative Adversarial Imitation Learning (GAIL) on three data domains: standard benchmarks, random state-action pairs, and egocentric scenarios. Simulated results suggest that simpler training methods (BC) are overall better than more complex training methods (RL), and that training samples with diverse agent-agent and agent-obstacle interactions are beneficial for reducing collisions in new scenarios.

Introduction

Simulating realistic crowd movements is crucial for computer graphics, architecture, and autonomous navigation. Recently, machine learning models have been used to imitate expert crowd behaviors by predicting an agent's next velocity based on its surroundings. However, most existing models are trained and tested in the same environment, merely varying the number of agents or their starting positions. This paper tackles the challenge of scenario generalization: how well an imitation model performs when dropped into a completely new environment with unseen obstacle configurations and agent destinations.

Methodology

The authors conduct a systematic evaluation crossing two representative training paradigms with three distinct data domains.

  1. Behavior Cloning (BC): A supervised learning approach that treats state-action pairs as independent samples, fitting a neural network regressor to map local observations to velocity decisions.
  2. Reinforcement Learning (RL): Specifically, Generative Adversarial Imitation Learning (GAIL), which trains a policy to match the "occupancy measure" (the distribution of states and actions) of the expert trajectories.

To standardize inputs, agents observe their surroundings via local range maps (distances to obstacles/agents), radial velocity maps, and local/global guidance velocities. The study utilizes three simulated data domains: Exocentric Standard (X) (classic benchmarks), Egocentric Representative (G) (4,000 diverse random scenarios), and Egocentric Random (R) (independent snapshots).

Key Results and Discussion

Comparison of RLA-G and BCA-G trajectories across standard scenarios, demonstrating the superior generalization of Behavior Cloning models.

Models were evaluated on Dynamic Time Warping (DTW), Agent-Agent collisions (AA), and Agent-Obstacle collisions (AO). The findings reveal several key insights:

  • Simpler is better for generalization: Behavior Cloning (BC) consistently generalized better to new scenarios than the more sophisticated RL (GAIL) approach.
  • The RL Overfitting Trap: RL struggles in scenario generalization because it relies on matching "occupancy measures" with the expert. This implicitly encodes the specific dynamics and obstacle layouts of the training environment into the policy. When the environment changes, the RL model's assumptions break down.
  • Diversity beats complexity: Training on a massive variety of small-scale, random interactions (Egocentric Representative/Random) proved more beneficial for reducing collisions in novel environments than training on a few highly complex, crowded scenarios (Exocentric Standard).

Visual analysis (as seen in the figure above) shows that while RL models (RLA-G) might have fewer collisions in some cases, it is often due to overly cautious behavior where agents wander and fail to reach their destinations. In contrast, BC models (BCA-G) maintain more direct and expert-like trajectories.

Conclusion

The study highlights that for robust crowd simulation in unseen environments, simpler training paradigms like Behavior Cloning combined with highly diverse training data are more effective than complex reinforcement learning methods. Future work will explore domain adaptation to further refine these models for specific real-world dynamics.

Summary updated

Abstract

Crowd simulation, the study of the movement of multiple agents in complex environments, presents a unique application domain for machine learning. One challenge in crowd simulation is to imitate the movement of expert agents in highly dense crowds. An imitation model could substitute an expert agent if the model behaves as good as the expert. This will bring many exciting applications. However, we believe no prior studies have considered the critical question of how training data and training methods affect imitators when these models are applied to novel scenarios. In this work, a general imitation model is represented by applying either the Behavior Cloning (BC) training method or a more sophisticated Generative Adversarial Imitation Learning (GAIL) method, on three typical types of data domains: standard benchmarks for evaluating crowd models, random sampling of state-action pairs, and egocentric scenarios that capture local interactions. Simulated results suggest that (i) simpler training methods are overall better than more complex training methods, (ii) training samples with diverse agent-agent and agent-obstacle interactions are beneficial for reducing collisions when the trained models are applied to new scenarios. We additionally evaluated our models in their ability to imitate real world crowd trajectories observed from surveillance videos. Our findings indicate that models trained on representative scenarios generalize to new, unseen situations observed in real human crowds.
BibTeX
@inproceedings{qiao2019scenario,
  author = {Gang Qiao and Honglu Zhou and Mubbasir Kapadia and Sejong Yoon and Vladimir Pavlovic},
  title = {{Scenario Generalization of Data-driven Imitation Models in Crowd Simulation}},
  booktitle = {MIG},
  year = {2019},
  doi = {10.1145/3359566.3360087},
}
Figure