Back to search

Clemson University

Training Physics-based Controllers for Articulated Characters with Deep Reinforcement Learning

Abstract

dc:description.abstract

In this thesis, two different applications are discussed for using machine learning techniques to train coordinated motion controllers in arbitrary characters in absence of motion capture data. The methods highlight the resourcefulness of physical simulations to generate synthetic and generic motion data that can be used to learn various targeted skills. First, we present an unsupervised method for learning loco-motion skills in virtual characters from a low dimensional latent space which captures the coordination between multiple joints. We use a technique called motor babble, wherein a character interacts with its environment by actuating its joints through uncoordinated, low-level (motor) excitation, resulting in a corpus of motion data from which a manifold latent space can be extracted. Using reinforcement learning, we then train the character to learn locomotion (such as walking or running) in the low-dimensional latent space instead of the full-dimensional joint action space. The thesis also presents an end-to-end automated framework for training physics-based characters to rhythmically dance to user-input songs. A generative adversarial network (GAN) architecture is proposed that learns to generate physically stable dance moves through repeated interactions with the environment. These moves are then used to construct a dance network that can be used for choreography. Using DRL, the character is then trained to perform these moves, without losing balance and rhythm, in the presence of physical forces such as gravity and friction.

Degree

thesis:*
Name thesis:degree_name
Master of Science (MS)
Level thesis:degree_level
Thesis
Discipline thesis:degree_discipline
School of Computing

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Biswas, Avishek
Contributors dc:contributor
  • Ioannis Karamouzas
  • Daljit S Dhillon

Subjects

dc:subject × 2

Identifiers

dc:identifier.*
Repository record dc:identifier
https://open.clemson.edu/all_theses/3518

Chain of custody

source
Harvested from
Clemson University
Base URL
open.clemson.edu/do/oai/
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Biswas, Avishek. Training Physics-based Controllers for Articulated Characters with Deep Reinforcement Learning. Thesis thesis, https://open.clemson.edu/all_theses/3518