Morphology agnostic multi-agent character control

Date

2025

Authors

Zhang, Rui

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

Crowd simulation plays a crucial role in various applications, from urban planning to virtual reality, by modeling realistic pedestrian behavior and interactions. Traditional approaches typically utilize simplified agent representation such as particles, whereas recent advancements have introduced fully physical character models in crowds, which relies on morphology-specific motion control, limiting their applicability to heterogeneous agents with diverse body structures and movement capabilities. This thesis introduces a morphology-agnostic multi-agent character control framework that integrates physics-based locomotion with hierarchical reinforcement learning. A low-level locomotion controller utilizes generalized goal conditioning to enable robust and adaptable movement across agents with different morphologies through parameter sharing, eliminating the need for predefined gait cycles or morphology-specific trajectory planning. A high-level navigation controller processes morphology-agnostic state observations and integrates visual attention sampling to improve decision-making. The navigation controller provides goal conditioning to the locomotion controller, guiding agents toward their target positions in dynamic environments. The proposed system improves generalizability in multi-agent settings by decoupling locomotion control from agent-specific kinematics while maintaining stability and responsiveness.

Description

Keywords

Crowd simulation, Reinforcement learning

Citation