← ArchivePaper2022

ControlVAE: Model-Based Learning of Generative Controllers for Physics-Based Characters

Heyuan Yao, Zhenhua Song, Baoquan Chen, Libin Liu

SIGGRAPH AsiaAcademic66 cites1 descendantMotion Synthesis

VAE-based framework learns a latent skill space from unstructured mocap and trains a generative physics-based motion controller.

Open the paper ↗Trace in the lineage graph

Abstract

In this paper, we introduce ControlVAE, a novel model-based framework for learning generative motion control policies based on variational autoencoders (VAE). Our framework can learn a rich and flexible latent representation of skills and a skill-conditioned generative control policy from a diverse set of unorganized motion sequences, which enables the generation of realistic human behaviors by sampling in the latent space and allows high-level control policies to reuse the learned skills to accomplish a variety of downstream tasks. In the training of ControlVAE, we employ a learnable world model to realize direct supervision of the latent space and the control policy. This world model effectively captures the unknown dynamics of the simulation system, enabling efficient model-based learning of high-level downstream tasks. We also learn a state-conditional prior distribution in the VAE-based generative control policy, which generates a skill embedding that outperforms the non-conditional priors in downstream tasks. We demonstrate the effectiveness of ControlVAE using a diverse set of tasks, which allows realistic and interactive control of the simulated characters.

How to read this

Category: Method: model-based VAE framework for physics-based motion control
Contributions: A VAE framework that learns a rich latent skill representation and a skill-conditioned generative control policy from unorganized motion sequences
A learnable world model that supervises the latent space and policy, enabling efficient model-based learning of downstream tasks
A state-conditional prior in the generative policy that yields skill embeddings outperforming non-conditional priors on downstream tasks
Context: Builds on latent-skill motion modeling such as Character Controllers Using Motion VAEs (Ling et al. 2020), adding a learned world model for direct latent/policy supervision in a physics-based setting.Builds on: Character Controllers Using Motion VAEs
Correctness: Premise is that a learned world model can stand in for unknown simulation dynamics to supervise the latent space; demonstrated across a diverse set of simulated tasks, so caveats are reliance on world-model fidelity and that results are interactive-simulation evaluations.
Clarity: The high-level VAE-plus-world-model story is followable; the model-based supervision and conditional prior reward a second pass.
How to read it: First pass for how the world model supervises skills versus a plain motion VAE; second pass on the training scheme and state-conditional prior if you plan to reuse or extend the controller.

Builds on

Character Controllers Using Motion VAEs 2020

Built upon by

MoConVQ: Unified Physics-Based Motion Control via Scalable Discrete Representations 2024

Related work

MuscleVAE: Model-Based Controllers of Muscle-Actuated Characters 2023 / SIGGRAPH Asia
Character Controllers Using Motion VAEs 2020 / SIGGRAPH
SFV: Reinforcement Learning of Physical Skills from Video 2018 / SIGGRAPH Asia
PADL: Language-Directed Physics-Based Character Control 2022 / SIGGRAPH Asia

Keywords

controlvae variational autoencoder physics-based character animation motion control deep reinforcement learning world model model-based learning skill embedding latent space state-conditional prior model predictive control supertrack mixture-of-experts pd controller locomotion adversarial discriminator evidence lower bound downstream tasks

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 →