← ArchivePaper2021

AMP: Adversarial Motion Priors for Stylized Physics-Based Character Control

Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, Angjoo Kanazawa

SIGGRAPHAcademic12 descendantsMotion Synthesis

Adversarial motion prior enabling physics-based characters to reproduce diverse motion styles from unstructured motion clip datasets.

Open the paper ↗Trace in the lineage graph

Abstract

Synthesizing graceful and life-like behaviors for physically simulated characters has been a fundamental challenge in computer animation. Data-driven methods that leverage motion tracking are a prominent class of techniques for producing high fidelity motions for a wide range of behaviors. However, the effectiveness of these tracking-based methods often hinges on carefully designed objective functions, and when applied to large and diverse motion datasets, these methods require significant additional machinery to select the appropriate motion for the character to track in a given scenario. In this work, we propose to obviate the need to manually design imitation objectives and mechanisms for motion selection by utilizing a fully automated approach based on adversarial imitation learning. High-level task objectives that the character should perform can be specified by relatively simple reward functions, while the low-level style of the character's behaviors can be specified by a dataset of unstructured motion clips, without any explicit clip selection or sequencing. For example, a character traversing an obstacle course might utilize a task-reward that only considers forward progress, while the dataset contains clips of relevant behaviors such as running, jumping, and rolling.

How to read this

Category: Method: adversarial motion prior for physics-based character control
Contributions: An adversarial imitation-learning prior (AMP) that lets simulated characters reproduce motion styles from an unstructured clip dataset without manual clip selection or sequencing
Separation of high-level task objectives (simple reward functions) from low-level style (the motion dataset), removing hand-designed imitation objectives
Context: Builds on tracking-based physics control such as DeepMimic, replacing per-clip tracking objectives with a GAN-style discriminator that scores motion naturalness.Builds on: DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills
Correctness: Demonstrated on physically simulated characters performing task-driven behaviors with learned style; note that adversarial training can be unstable and prone to mode collapse, and style fidelity depends on the coverage of the provided clip dataset.
Clarity: Readable conceptually; a second pass is needed for the discriminator formulation and the reward-combination details.
How to read it: First pass for the task-reward vs. style-prior split (the core idea); second pass on the discriminator objective and training setup if you plan to apply it to your own characters.

Builds on

DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills 2018

Built upon by

Related work

Synthesizing Physical Character-Scene Interactions 2023 / SIGGRAPH
DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills 2018 / SIGGRAPH
DReCon: Data-Driven Responsive Control of Physics-Based Characters 2019 / TOG
ASE: Large-Scale Reusable Adversarial Skill Embeddings for Physically Simulated Characters 2022 / SIGGRAPH

Keywords

adversarial motion prior physics-based character control reinforcement learning generative adversarial imitation learning adversarial discriminator motion style transfer goal-conditioned reinforcement learning proximal policy optimization least-squares discriminator gradient penalty unstructured motion clips reference state initialization early termination motion prior pd controllers locomotion

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 →