Deep Imitation Learning for Complex Manipulation Task from Virtual Reality Teleoperation

Intelligent Robotics Seminar: Paper Review & Implementation, Universität Hamburg

This project presents a comprehensive review and analysis of mastering robot skills through imitation learning—a paradigm that has gained significant traction for its ability to facilitate complex demonstrations without extensive professional intervention. The study examines how specialized algorithms can be optimized to address the challenges of policy acquisition, specifically focusing on the mapping from raw pixels to robotic manipulation actions.

The work explores the underlying methodologies of using consumer-grade Virtual Reality (VR) headsets and hand-tracking hardware to naturally teleoperate robots. By analyzing deep neural network policies learned through imitation, the review evaluates how pixels-to-actions mapping demonstrates acquired skills in high-dimensional environments, such as table tennis.

Finally, we evaluate the reviewed imitation learning approaches through contextual video analysis, comparing them against traditional teleoperation and validating the proposed frameworks for robot skill acquisition. This review serves to synthesize current advancements in embodied AI and propose a robust foundation for future skill acquisition models.