Keywords

Robotic Motion Planning, Dexterous Manipulation, Manifold Learning, Trajectory Optimization, Generative Artificial Intelligence, Latent Space based Motion Planning.

Abstract

In modern robotics, effectively computing optimal robotic control policies under dynamically varying environmental constraints poses substantial challenges and remains a unique research endeavor. To compute efficient and safe robotic motion planning while achieving collective objectives, robotic agents have to satisfy two criteria. First, the working robotic entity must be capable of handling a non-stationary working environment with dynamic obstacles and system constraints. Second, to ensure the real-time response, the robotic agent has to compute an effective control policy that meets a real-time learning performance. Despite significant advancements in motion planning with the appearance of extensive computing resources and advanced deep learning algorithms, several formidable challenges remain in this robotic control design. In this dissertation, we develop techniques focusing on generalized methodology of robotic motion planning with efficiency and scalability that can quickly adapt to variable targets, dynamic obstacles, multiple robotic arm dynamics and successfully complete many intricate tasks related to non-stationary environments. In particular, we first aim at creating a unified framework for dexterous robotic manipulation that can handle dynamical environmental constraints such as dynamic obstacles, dynamic targets and multi-robot coordination without having a complete access to explicit environment models. Secondly, we propose computationally efficient and scalable real-time robotic control algorithms by integrating lower-dimensional manifold representation learning techniques with deep generative artificial intelligence based learning paradigm. Along with handling non-linear and unpredictable system constraints, we solve challenges with re-synthesizing robotic commands based on real-time object positions under brief reaction times, which ensures time optimization of generated robot motion. Finally, we propose a novel cross-embodiment robotic manipulation algorithm via bidirectional subspace alignment, cycle consistency and human behavior transformer inspired by robotic learning from expert demonstration to address the scarcity of paired cross-embodiment datasets and the impediment of designing intricate controllers. We develop an innovative methodology to address imbalanced datasets from heterogeneous domains and develop motion planning techniques through knowledge distillation. We observe that our findings exhibit significant implications for the future design of intelligent and autonomous robots, enhancing their capability to tackle more intricate robotic manipulation tasks with increased efficiency and safety.

Completion Date

2025

Semester

Spring

Committee Chair

Lin, Mingjie

Degree

Doctor of Philosophy (Ph.D.)

College

College of Engineering and Computer Science

Department

Electrical and Computer Engineering

Identifier

DP0029283

Document Type

Dissertation/Thesis

Campus Location

Orlando (Main) Campus

Download

Share

COinS