← ArchivePaper2024

ReGAIL: Toward Agile Character Control From a Single Reference Motion

Paul Marius Boursin, Yannis Kedadry, Victor B. Zordan, Paul G. Kry, Marie-Paule Cani

Relative-feature modifications to GAIL's discriminator observation enable agile physics-based character control from a single motion cycle.

Open the paper ↗Trace in the lineage graph

Abstract

We present an approach for training "agile" character control policies, able to produce a wide variety of motor skills from a single reference motion cycle. Our technique builds off of generative adversarial imitation learning (GAIL), with a key novelty of our approach being to provide modification to the observation map in order to improve agility and robustness. Namely, to support more agile behavior, we adjust the value measurements of the training discriminator through relative features - hence the name ReGAIL. Our state observations include both task relevant relative velocities and poses, as well as relative goal deviation information. In addition, to increase robustness of the resulting gaits, servo gains and damping values are included as part of the policy action to let the controller learn how to best combine tension and relaxation during motion. From a policy informed by a single reference motion, our resulting agent is able to maneuver as needed, at runtime, from walking forward to walking backward or sideways, turning and stepping nimbly. We demonstrate our approach for a humanoid and a quadruped, on both flat and sloped terrains, as well as provide ablation studies to validate the design choices of our framework.

How to read this

Category: Method: physics-based character control via imitation learning
Contributions: ReGAIL, training agile control policies from a single reference motion cycle
A relative-feature modification to the GAIL discriminator's observation map for agility and robustness
Including servo gains and damping in the policy action so the controller learns to combine tension and relaxation
Context: A physics-based control method in the adversarial-imitation lineage, building on AMP: Adversarial Motion Priors (Peng 2021) and GAIL.Builds on: AMP: Adversarial Motion Priors for Stylized Physics-Based Character Control
Correctness: Demonstrated on a humanoid and a quadruped across flat and sloped terrain with ablation studies; results come from a single reference cycle per character, so the variety of emergent skills is bounded by what that cycle and the simulation support.
Clarity: Accessible if familiar with RL/GAIL; a first pass conveys the relative-feature idea, a second pass for the observation and action design.
How to read it: First pass for the single-reference premise and the relative-feature trick; second pass on the discriminator observation map and the gain/damping action if you train policies yourself.

Builds on

AMP: Adversarial Motion Priors for Stylized Physics-Based Character Control 2021

Built upon by

Nothing yet.

Related work

AMP: Adversarial Motion Priors for Stylized Physics-Based Character Control 2021 / SIGGRAPH
Near-Optimal Character Animation with Continuous Control 2007 / SIGGRAPH
UniCon: Universal Neural Controller for Physics-Based Character Motion 2020 / arXiv
Physics-Based Motion Retargeting from Sparse Inputs 2023 / SCA

Keywords

reinforcement learning character animation motion control gail generative adversarial imitation learning locomotion physically-based simulation pd control servo gains relative observations goal-driven animation agile character control motion retargeting imitation learning

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 →