자료유형  

 학위논문

Control Number  

0017164867

International Standard Book Number  

9798346394648

Dewey Decimal Classification Number  

620

Main Entry-Personal Name  

Jiang, Yifeng.

Publication, Distribution, etc. (Imprint  

[S.l.] : Stanford University., 2024

Publication, Distribution, etc. (Imprint  

Ann Arbor : ProQuest Dissertations & Theses, 2024

Physical Description  

159 p.

General Note  

Source: Dissertations Abstracts International, Volume: 86-05, Section: B.

General Note  

Advisor: Liu, Karen.

Dissertation Note  

Thesis (Ph.D.)--Stanford University, 2024.

Summary, Etc.  

요약Creating digital twins of humans and robots in physics simulations has widespread applications in both digital and real-world domains, including gaming, 3D content generation, virtual and augmented reality, robot motor skill learning, synthetic data generation for large-scale AI model training, and computational healthcare. Unlike passive objects, physics simulations of virtual human-like characters and robots as active agents are closely intertwined with control mechanisms. This interplay presents unique challenges for existing general-purpose physics simulators, such as balancing accuracy and detail with the ease of control, and ensuring flexibility for customization and identifiability while maintaining physical correctness, even with limited observational data.The capability of deep neural networks to compress large amounts of data into compact representations o↵ers promising opportunities for creating next-generation simulations. This thesis explores the integration of deep learning techniques with simulation-based methods for simulating human-like characters and robots. The central idea is that controller design goals can o↵er critical insights into simulation design, guiding the e↵ective use of deep learning to automate customization, accelerate performance, and enable simple, standard control algorithms to synthesize diverse and realistic physical behaviors e↵ectively.The thesis begins with applications in robotics, focusing on learning data-driven simulators that better match unseen, novel dynamics compared to conventional non-learning simulation models, thereby minimizing the dynamics discrepancy between controller training and deployment. We investigate hybrid simulation models and new approaches to simulator learning using adversarial reinforcement learning. The focus then shifts to digital humans, where we learn biomechanics constraints for simulated characters. We propose a transform between spaces of simulation states that allows for faster simulation and easier control while maintaining biomechanical validity. Next, we introduce a novel simulation framework based on projective dynamics and generative models of human movements. In this framework, kinematic generative models, trained without explicitly considering physics, serSve as control energy in the simulation of human skeletons, thus bypassing the need for controller training with a black-box general-purpose simulator in the loop and fully leveraging the scalability and robustness of kinematic models. The thesis concludes with a coste↵ective solution for human motion capture, capable of providing large-scale datasets to meet the data requirements of the aforementioned digital human applications.

Subject Added Entry-Topical Term  

Robots.

Subject Added Entry-Topical Term  

Kinematics.

Subject Added Entry-Topical Term  

Physics.

Subject Added Entry-Topical Term  

Elbow.

Subject Added Entry-Topical Term  

Robotics.

Added Entry-Corporate Name  

Stanford University.

Host Item Entry  

Dissertations Abstracts International. 86-05B.

Electronic Location and Access  

로그인을 한후 보실 수 있는 자료입니다.

Control Number  

joongbu:656193