PFNN – a neural network in Metro Exodus

What the talk is about

Maksym Hanenko of 4A Games explains how Metro Exodus used the PFNN neural network, built around a phase function, to animate character locomotion. Instead of an animation state graph with its combinatorial explosion of transitions, the network is trained on recorded actor movement and predicts the next pose each frame from the current pose and the desired path. The speaker describes his contributions: data visualization, automatic detection of anomalies in the dataset, footstep labeling with pressure-sensor insoles, and steering NPCs with an imaginary joystick that corrects itself based on the error against the desired trajectory. Moving the dataset to the GPU and changing the phase representation cut an epoch from 15 minutes to 30 seconds. In the Q&A he covers the architecture and development cost. Useful for programmers and animators interested in ML for game animation.