On the Move: Towards Realistic and Controllable Human Motion Generation

[eng] In recent years, we have witnessed the arrival of algorithms capable of understanding human behavior and generating the motion that drives virtual human-like avatars or robots. This ability, known as human motion generation, has emerged thanks to major advances in deep learning, parameterized...

Descripción completa

Detalles Bibliográficos
Autor: Barquero Garcia, German
Tipo de recurso: tesis doctoral
Estado:Versión publicada
Fecha de publicación:2026
País:España
Institución:Universidad de Barcelona
Repositorio:Dipòsit Digital de la UB
OAI Identifier:oai:dnet:ubarcelona__::690697de5d123e03061f93bfde8d94f1
Acceso en línea:https://hdl.handle.net/2445/228711
https://hdl.handle.net/10803/697193
Access Level:acceso abierto
Palabra clave:Intel·ligència artificial
Aprenentatge profund
Mecànica humana
Artificial intelligence
Deep learning (Machine learning)
Human mechanics
Descripción
Sumario:[eng] In recent years, we have witnessed the arrival of algorithms capable of understanding human behavior and generating the motion that drives virtual human-like avatars or robots. This ability, known as human motion generation, has emerged thanks to major advances in deep learning, parameterized human body models, and large-scale motion capture datasets. Together, these developments have enabled the synthesis of lifelike movement from minimal or abstract input such as natural language descriptions, or short clips of prior motion. These generative capabilities are unlocking new applications in animation, extended reality, robotics, and behavior-aware autonomous systems, helping to design digital humans or humanoids that both look and behave in a human-like way. Such advances represent an essential step towards developing fully immersive, human-centric experiences. This thesis focuses on enhancing human motion generation along two key axes: realism and control. Specifically, I make progress in three important topics. First, in human motion prediction, I show that the prevailing focus of maximizing coordinate-level diversity encourages unrealistic predictions. I overturn this status quo by proposing a model that delivers behavioral diversity and state-of-the-art realism. Second, I improve the controllability of motion generation models by enabling users to specify both the description and duration of consecutive actions within arbitrarily long motion sequences. For this, I propose a new technique that eliminates the need for post-processing, and promotes smooth and realistic transitions between actions. Finally, I address the challenge of real-time full-body motion synthesis from head-mounted sensors and vision-based hand-tracking inputs, which are often noisy and unreliable. In particular, I present the first method able to preserve motion continuity and realism through signal losses. I complement it with the release of the first dataset with paired headset-captured tracking and ground-truth motion capture during real virtual-reality interactions. I use it to reveal the performance gap when deploying existing methods trained on synthetic data in real-life conditions. In addition to these three core contributions, this thesis also introduces novel evaluation metrics that improve the way motion quality is assessed. I propose two smoothness metrics that correlate with perceptual quality in motion prediction, and jerk-based transition metrics that quantify motion discontinuities during action transitions. Collectively, these contributions push the boundaries of human motion generation in realism and controllability and provide a toolkit for next-generation systems that must generate lifelike human movement on demand.