Graduation Semester and Year
Summer 2026
Language
English
Document Type
Dissertation
Degree Name
Doctor of Philosophy in Mechanical Engineering
Department
Mechanical and Aerospace Engineering
First Advisor
Panos S. Shiakolas
Second Advisor
Animesh Chakravarthy
Third Advisor
Haying Huang
Fourth Advisor
Seiichi Nomura
Fifth Advisor
Shouyi Wang
Abstract
Anthropomorphic Robot Hands (ARHs) are versatile end-effectors in assistive systems, designed to mimic human hand dexterity for complex manipulation tasks. Their ability to interact with diverse objects, from rigid tools to delicate, deformable items, makes them invaluable for supporting daily activities. While significant progress has been made in grasping rigid objects, grasping deformable objects with multi-fingered hands like ARHs remains an open challenge. This research investigates and develops learning-based grasp synthesis methods for deformable objects using ARHs. The research begins with the development of an assistive system for human-guided grasp synthesis of rigid objects, integrating a vibrotactile-enabled flex-glove, virtual simulation, and an underactuated ARH1 (InMoov hand). The architecture was evaluated with two control methods, using a synthesized grasp and under real-time continuous control, and the evaluation validated the system for human-guided grasp synthesis. The synthesized grasps were then evaluated through rigid-body simulation in CoppeliaSim using established grasp quality indices, which quantified the achievable grasp quality and identified the limitations imposed by the underactuated 1-DOF thumb. A second-generation hand, ARH2, was developed with a fully actuated 4-DOF thumb to enhance grasping capability. An affordable and personalizable carpometacarpal (CMC) joint tracking device and a kinematic algorithm were developed to map human thumb motion to the kinematically different robot thumb. This combination enabled dexterous grasp synthesis across a range of object shapes and sizes, including tracked thumb motion with the other fingers. A custom teleoperation system was interfaced with NVIDIA Isaac Sim to interact with volumetric deformable objects through human-guided demonstrations for ARH2 grasp synthesis. The first investigation employs human-guided demonstrations and contact-driven synthetic augmentation to build an expanded demonstration dataset for Behavioral Cloning (BC). Fifty-four successful human-guided demonstrations were expanded with 200 synthetic grasps to form a 254-demonstration dataset. A Transformer BC policy trained on this dataset was evaluated offline and in Isaac Sim on held-out cylindrical and cuboid deformable objects, achieving grasp execution rates of 68% and 56%, respectively. The second investigation develops an optimization-based geometric grasp generation pipeline for ARH2, coupled with reinforcement learning (RL) for deformable grasp modulation. Candidate grasps for cylinder and cuboid objects are generated with an ARH2-adapted formulation that includes a thumb-opposition energy term, then filtered by physics-based validation in Isaac Sim. An RL policy trained with Proximal Policy Optimization learns bounded residual corrections around the validated grasps under a settle, hold, and robustness curriculum. The adapted generator reduced mean optimization energy by 35% relative to the baseline, validation improved the yield of usable grasps by a relative 25%, and the trained policy settled in 72% of evaluated episodes, held under gravity in 64%, and survived a full gravity-vector rotation in 58%. Overall, this research extends learning-based grasp synthesis for ARHs on rigid and deformable objects through BC with synthetic augmentation and through geometric grasp generation with RL modulation. The combined results demonstrate how complementary machine learning methods can advance scalable and robust ARH deformable grasping pipelines.
Keywords
Anthropomorphic robotic hand, Deformable object grasping, Grasp Generation, Behavioral cloning, Reinforcement learning, Imitation learning, Teleoperation, Human-guided Grasp Synthesis, Assistive robotics, Physics-based Simulation
Disciplines
Acoustics, Dynamics, and Controls | Electro-Mechanical Systems | Human-Computer Interaction | Robotics
License
This work is licensed under a Creative Commons Attribution-NonCommercial-No Derivative Works 4.0 International License.
Recommended Citation
Abdul Rahaman, Abdul Hafiz, "INVESTIGATION AND DEVELOPMENT OF LEARNING-BASED GRASP SYNTHESIS METHODS FOR RIGID AND DEFORMABLE OBJECTS USING ANTHROPOMORPHIC ROBOT HANDS" (2026). Mechanical and Aerospace Engineering Dissertations. 9.
https://mavmatrix.uta.edu/mechaerospace_dissertations2/9
Included in
Acoustics, Dynamics, and Controls Commons, Electro-Mechanical Systems Commons, Human-Computer Interaction Commons, Robotics Commons