← ArchivePaper2020

Learned Motion Matching

Daniel Holden, Oussama Kanoun, Maksym Perepichka, Tiberiu Popa

SIGGRAPHUbisoft5 descendantsMotion Synthesis

Compresses motion matching into neural networks, keeping its controllability with a fraction of the memory.

Open the paper ↗Trace in the lineage graph

Abstract

This paper introduces Learned Motion Matching, a neural-network-based alternative to the Motion Matching algorithm that preserves its quality, control, and quick iteration time while achieving the scalability and low memory usage of generative models. The Motion Matching algorithm is broken into three stages, Projection, Stepping, and Decompression, each replaced by a specialized neural network, the Projector, Stepper, and Decompressor, together with an autoencoder-like Compressor that discovers latent variables. The resulting model removes the need to store the matching and animation databases in memory, so memory no longer scales linearly with the amount of animation data. The method is demonstrated on locomotion, rough terrain, chair interaction, character interactions, and quadruped characters, and a user study in a AAA production found participants could barely distinguish it from basic Motion Matching.

How to read this

Category: Method: a neural reformulation of Motion Matching
Contributions: Learned Motion Matching, replacing Motion Matching's Projection, Stepping, and Decompression stages with a Projector, Stepper, and Decompressor network.
An autoencoder-like Compressor that discovers latent variables so the matching and animation databases need not be stored in memory.
Memory that no longer scales linearly with the animation data while preserving the quality, control, and iteration speed of Motion Matching.
Context: Directly builds on Clavet's 'Motion Matching and The Road to Next-Gen Animation' and on neural character control such as Holden et al.'s Phase-Functioned Neural Networks.Builds on: Motion Matching and The Road to Next-Gen Animation · Phase-Functioned Neural Networks for Character Control
Correctness: Demonstrated across locomotion, rough terrain, chair interaction, character interactions, and quadrupeds, with a AAA-production user study reporting it was hard to distinguish from Motion Matching; as with learned models, behavior outside the training database is the main thing to watch.
Clarity: Accessible if you already know Motion Matching; a first pass conveys the three-network decomposition, a second pass clarifies training of each stage.
How to read it: Anchor on the original Motion Matching stages, then map each to its network; a second pass on the Compressor and training setup is worth it if you plan to deploy or reproduce it.

Builds on

Built upon by

Related work

Neural State Machine for Character-Scene Interactions 2019 / SIGGRAPH Asia
Phase-Functioned Neural Networks for Character Control 2017 / SIGGRAPH
DReCon: Data-Driven Responsive Control of Physics-Based Characters 2019 / TOG
Learning Robust and Scalable Motion Matching with Lipschitz Continuity and Sparse Mixture of Experts 2023 / MIG

Keywords

motion matching learned motion matching neural networks generative models character control decompressor compressor stepper projector nearest neighbor search autoregressive latent variables motion capture locomotion inertialization blending forward kinematics data augmentation quadruped

This page summarises the entry and links to its original source. The archive never hosts or redistributes the publication itself.Show it in the full archive list →