Phase-Functioned Neural Networks for Character Control

How to read this

Category

Method: a neural network architecture for real-time character control

Contributions

• The Phase-Functioned Neural Network, whose weights are regenerated each frame by a cyclic phase function (a cubic Catmull-Rom spline over the motion-cycle phase)
• Real-time, terrain-adaptive locomotion (walking, running, jumping, climbing, crouching) from user controls, prior state, and surrounding terrain geometry
• A compact, fast model that avoids autoregressive RNN dying-out and blending artifacts while training on motion fitted to large heightmap databases

Context

Bridges data-driven control approaches such as Clavet's Motion Matching and the deep-learning motion synthesis line (Holden et al.), introducing explicit phase conditioning for character controllers.Builds on: Motion Matching and The Road to Next-Gen Animation · A Deep Learning Framework for Character Motion Synthesis and Editing

Correctness

Demonstrated on locomotion with explicit cyclic phase; readers should remember the phase assumption suits cyclic motion well, that preparing training data requires fitting captures to terrain heightmaps, and that non-cyclic actions are a less natural fit for this parameterization.

Clarity

Well written and influential; a first pass conveys the phase-function idea clearly, a second pass covers the data preparation and network details.

How to read it

Focus on the phase-function weight-generation idea and why it beats autoregressive blending; a second pass on data fitting and the spline parameterization pays off given how foundational this work is.