Yin, Qi
(2024)
Exploring the Reality Gap: Deep Reinforcement Learning for Training a 6DOF Robotic Arm to Grasp Target Box.
Master's Thesis, University of Pittsburgh.
(Unpublished)
Abstract
This thesis investigates a significant challenge in the application of reinforcement learning to robotics: the "reality gap", which refers to the differences in behavior and performance exhibited by robots trained in simulated environments when applied to real-world scenarios. The study focuses on a robotic arm trained to grasp a target box through reinforcement learning, thoroughly analyzing the process of modeling the robotic arm, training it using reinforcement learning techniques, and transferring the learned behaviors to real-world applications. The research underscores the importance of accurately capturing physical properties such as mass, friction, and inertia in simulations and discusses the complexities involved in modeling actuators and control systems for tasks requiring precise manipulation.
Through both qualitative and quantitative analyses, this study examines the discrepancies between simulation and reality, identifying key factors contributing to the reality gap. The research provides a detailed comparison using visual imagery and joint angle data to measure the performance differences between simulated training and real-world execution of the robotic arm. Additionally, the study explores the application of the Deep Deterministic Policy Gradient (DDPG) algorithm in training, highlighting its effectiveness and the challenges faced in translating simulation achievements into real-world operations.
Physics engines face significant challenges when simulating complex contact forces, such as friction and collision forces, leading to substantial discrepancies between computational models and real-world scenarios. The research findings further highlight the limitations of current simulation models, particularly pointing out that limitations in modeling and calculating complex contact forces are among the key factors affecting the transferability of simulated training to practical tasks. The thesis concludes with a discussion on potential strategies to narrow the reality gap, suggesting future research directions aimed at enhancing the accuracy of simulation models and improving the real-world applicability of robotic training. This work aims to bridge the gap between theoretical models and their practical implementation, thereby improving the efficiency and reliability of Reinforcement Learning in real-world applications.
Share
| Citation/Export: |
|
| Social Networking: |
|
Details
| Item Type: |
University of Pittsburgh ETD
|
| Status: |
Unpublished |
| Creators/Authors: |
|
| ETD Committee: |
|
| Date: |
6 September 2024 |
| Date Type: |
Publication |
| Defense Date: |
19 April 2024 |
| Approval Date: |
6 September 2024 |
| Submission Date: |
4 May 2024 |
| Access Restriction: |
No restriction; Release the ETD for access worldwide immediately. |
| Number of Pages: |
77 |
| Institution: |
University of Pittsburgh |
| Schools and Programs: |
Swanson School of Engineering > Mechanical Engineering |
| Degree: |
MS - Master of Science |
| Thesis Type: |
Master's Thesis |
| Refereed: |
Yes |
| Uncontrolled Keywords: |
Reinforcment learning, Sim2Real, Reality gap, Robotic manipulation |
| Date Deposited: |
06 Sep 2024 19:54 |
| Last Modified: |
06 Sep 2024 19:54 |
| URI: |
http://d-scholarship.pitt.edu/id/eprint/46386 |
Metrics
Monthly Views for the past 3 years
Plum Analytics
Actions (login required)
 |
View Item |
![[feed]](/images/feed-icon-32x32.png){kind=icon}