Movement-primitive-based imitation learning for robotic manipulation in domestic environment

• Main Author: Xu, Ziqi
• Other Authors: Wen Bihan
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Engineering
• Online Access: https://hdl.handle.net/10356/181818
• Institution: Nanyang Technological University

With the rapid advancement of robotics and artificial intelligence, robots are increasingly being integrated into domestic environments to assist with daily tasks. However, enabling robots to perform complex manipulation tasks autonomously in unstructured and dynamic environments remains a major challenge. This dissertation presents the development of a general-purpose robotic manipulation platform capable of replicating complex human skills using imitation learning techniques. The platform utilizes Dynamic Movement Primitives (DMPs) and Probabilistic Movement Primitives (ProMPs) to model and generalize nonlinear trajectories, enabling the robot to adapt to diverse task requirements in real-time. The proposed system focuses on two key aspects of robotic manipulation: goal-oriented and trajectory-oriented tasks. Goal-oriented tasks involve actions such as picking and placing objects, while trajectory-oriented tasks require more intricate movements like folding or stirring. By integrating DMPs and ProMPs, the platform ensures that the robot can handle a wide variety of manipulation tasks with generality and adaptability. Imitation learning serves as the primary learning method, allowing the robot to acquire skills demonstrated by human experts. Experimental results show that the platform can generalize learned behaviors to new tasks and environments with minimal retraining, making it highly suitable for domestic applications. This dissertation makes contributions to the field of robotic manipulation by developing a robust framework that combines imitation learning with movement primitives. The system's ability to handle nonlinear, high-dimensional tasks in real-time displays its potential for broader applications in both domestic and industrial settings. Future work includes optimizing the system for more complex environments, integrating LLMs and exploring its application in other fields where human-robot collaboration is critical.