Agenda

This session presents a production-proven Neural Simulation system for skeletal physics, as seen on dozens of characters in the action game FateTrigger. The presenters explain how they overcame the notorious performance bottleneck of real-time secondary physics simulations (hair, cloth, accessories) for a large cast of characters. Their deep learning solution delivers high-fidelity, stable physics for a fraction of the cost—just 0.02-0.13ms on a single CPU core—making complex simulations viable while maintaining smooth, high-framerate gameplay. The talk explores the entire production pipeline, from its meticulously designed network architecture with quantization-aware training, to its high-performance custom C++ inference engine and a dynamic, load-balanced LOD system. It demonstrates that neural physics is no longer a research concept but a powerful solution for achieving superior performance while scaling to a higher volume of dynamic character simulations.

Takeaway

Attendees gain a practical blueprint for building their own high-performance neural physics system. This includes specific techniques for creating an automated data pipeline, designing a versatile network architecture for temporal dynamics, and deploying a highly optimized C++ inference engine for real-time performance in a production environment.

Intended Audience

This session is for graphics programmers, engine developers, AI programmers, and technical artists. A solid understanding of 3D math and game engine architecture is recommended. Basic familiarity with machine learning concepts and traditional character physics systems will be beneficial for a deeper understanding of the material.