← ArchivePaper2020

MoGlow: Probabilistic and Controllable Motion Synthesis Using Normalising Flows

Gustav Eje Henter, Simon Alexanderson, Jonas Beskow

SIGGRAPH AsiaAcademic109 citesMotion Synthesis

Autoregressive normalising-flow model produces diverse, controllable locomotion sequences trained with exact maximum likelihood.

Open the paper ↗Trace in the lineage graph

Abstract

Data-driven modelling and synthesis of motion is an active research area with applications that include animation, games, and social robotics. This paper introduces a new class of probabilistic, generative, and controllable motion-data models based on normalising flows. Models of this kind can describe highly complex distributions, yet can be trained efficiently using exact maximum likelihood, unlike GANs or VAEs. Our proposed model is autoregressive and uses LSTMs to enable arbitrarily long time-dependencies. Importantly, is is also causal, meaning that each pose in the output sequence is generated without access to poses or control inputs from future time steps; this absence of algorithmic latency is important for interactive applications with real-time motion control. The approach can in principle be applied to any type of motion since it does not make restrictive, task-specific assumptions regarding the motion or the character morphology. We evaluate the models on motion-capture datasets of human and quadruped locomotion. Objective and subjective results show that randomly-sampled motion from the proposed method outperforms task-agnostic baselines and attains a motion quality close to recorded motion capture.

How to read this

Category: Method: a normalising-flow model for probabilistic, controllable motion synthesis
Contributions: MoGlow, a probabilistic generative motion model based on normalising flows, trainable by exact maximum likelihood (unlike GANs or VAEs).
An autoregressive design using LSTMs for arbitrarily long time-dependencies, and a causal formulation with no algorithmic latency for real-time control.
A task-agnostic approach making no restrictive assumptions about the motion or character morphology, evaluated on human and quadruped locomotion.
Context: Relates to neural character control such as Holden et al.'s Phase-Functioned Neural Networks, offering a flow-based probabilistic alternative to GAN/VAE motion models.Builds on: Phase-Functioned Neural Networks for Character Control
Correctness: Objective and subjective evaluations on motion-capture locomotion data report sampled motion outperforming task-agnostic baselines; being task-agnostic and data-driven, quality and diversity remain bounded by the training motion, and the abstract's quality claim is truncated.
Clarity: Accessible in framing but requires flow-model background; a first pass conveys the probabilistic, causal, controllable design, a second pass clarifies the normalising-flow formulation.
How to read it: Focus first on why exact-likelihood flows and causality matter for interactive control; a second pass on the flow and autoregressive conditioning is worth it if you want to train or sample motions.

Builds on

Phase-Functioned Neural Networks for Character Control 2017

Built upon by

Nothing yet.

Related work

Local Motion Phases for Learning Multi-Contact Character Movements 2020 / SIGGRAPH
Learned Motion Matching 2020 / SIGGRAPH
Robust Motion In-Betweening 2020 / SIGGRAPH
Multi-Objective Adversarial Gesture Generation 2019 / MIG

Keywords

normalising flows glow probabilistic motion synthesis controllable motion generation autoregressive model lstm maximum likelihood affine coupling layer actnorm locomotion synthesis data dropout motion capture quadruped locomotion causal model foot sliding generative models

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 →