Key points
- PFNN modulates network weights by a cyclic motion phase, producing smooth real-time locomotion with terrain interaction.
- It is the named root of the Mode-Adaptive Network, Neural State Machine and DeepPhase line.
- It anchors the widely used AI4Animation teaching framework.
Summary
Holden, Komura and Saito (2017) introduced Phase-Functioned Neural Networks, in which the weights of a small network are computed as a function of a cyclic motion phase variable by a second "gating" network. This allows the controller to smoothly interpolate movement styles across the locomotion cycle and terrain types, producing responsive real-time character locomotion with terrain adaptation. Published at SIGGRAPH 2017, PFNN is the named origin of the phase-based neural locomotion family and anchors the AI4Animation framework widely used in academic and industry teaching. It is the upstream paper to the held DeepPhase entry and distinct from Holden's 2016 motion-synthesis paper.
Related items
- DeepPhase: Periodic Autoencoders for Learning Motion Phase Manifolds
- A deep learning framework for character motion synthesis and editing (SIGGRAPH 2016 / SIGGRAPH 2026 Test-of-Time award)
Source
Source: ACM Transactions on Graphics (SIGGRAPH 2017) ↗ (Research)
Cite this item
Holden, D., Komura, T. & Saito, J. (2017). ‘Phase-Functioned Neural Networks for Character Control (PFNN)’, ACM Transactions on Graphics (SIGGRAPH 2017). Available at: https://dl.acm.org/doi/10.1145/3072959.3073663
Your reference manager can also read this page directly: with the Zotero (or Mendeley) browser connector installed, save it straight to your library. Whole-collection exports: RIS, BibTeX, CSL-JSON.