Benchmarking ML based navigation systems in video games

This project develops and benchmarks novel offline training algorithms for AI navigation in gaming environments. As such, it is intended to be a self-contained project.
Traditional video game navigation heavily relies on navigation meshes (navmeshes) for pathfinding. Navmeshes offer a simplified representation of complex environments but face limitations in portraying nuanced navigation abilities, such as climbing or jumping. These abilities often necessitate additional constructs like navigation links (navlinks), which, while functional, can be unwieldy and less adaptable to dynamic game elements. Additionally, navmeshes can struggle to scale effectively in intricately detailed game environments, often requiring a compromise between accuracy and computational efficiency.
The evolution of game AI has seen a shift towards online learning methods like Reinforcement Learning (RL) and algorithms such as the Soft Actor-Critic (SAC). SAC, with its efficiency in handling continuous action spaces, has shown potential in navigating complex environments. However, the reliance on extensive simulations for training poses considerable challenges, including long iteration times and high resource demands.
In this context, offline training methods like Behavior Cloning and Goal-Conditioned Behavioral Cloning (GCBC) emerge as promising alternatives. Offline training, or batch RL, leverages pre-collected data for AI training, thereby circumventing the need for ongoing interaction with the environment. This approach significantly streamlines the development process by reducing iteration times.
Our aim is to improve AI navigation’s efficiency and effectiveness in dynamic game scenarios by leveraging offline training methods like BC and GCBC. We propose a dual-stage benchmarking process.
Initially, we’ll use a basic prototype environment in Godot for rapid algorithm iteration and refinement. Godot’s simplicity aids in early-stage algorithm tuning. Following this, we’ll escalate testing to a complex game currently under development at Ubisoft, offering a real-world application scenario. This step stress-tests the algorithms in a sophisticated, large-scale game setting, evaluating their robustness and scalability.
This two-pronged approach allows for rigorous testing from simple to complex environments, ensuring the algorithms’ applicability in diverse gaming contexts. Our research intends to significantly advance AI navigation in gaming, providing insights applicable to AI in interactive environments.

Faculty Supervisor:

Amir-massoud Farahmand;Sheila McIlraith

Student:

Partner:

Ubisoft Toronto

Discipline:

Computer science

Sector:

Information and cultural industries

University:

University of Toronto

Program: