Master's Thesis

Tight clearances have always been a posing challenge while performing industrial assembly tasks. In many cases, robotic systems have positional errors larger than the clearance between assembly parts of the task. A Peg-In-Hole setup, with clearance 0.3 mm, is utilized in this thesis to conduct experiments of different approaches: Autonomous, Shared, Offline LfD, and Online LfD. We measure the Success Rate (SR) and Completion Time (CT) as performance indicators in all experiments. AUTO: The Peg picking part is done autonomously using (Deep Learning and/or Classical) CV techniques to detect the random peg position from the starting area. The Peg insertion part is done using generated masks to detect the hole position; thus, complete the Peg-In-Hole task. SHARED: The task is performed autonomously until a termination condition occurs; whether the autonomous searching time reaches a timeout or the user decides to take over the control. The user then completes the task via Bi-lateral Tele-operation system. Learning from Demonstration (LfD): User teleoperated demonstrations (or trials) are utilized to learn a Gaussian Mixture Model (GMM) model encoding the task trajectory in 3D space. Gaussian Mixture Regression (GMR) is used to infer a learned trajectory for performing the Peg-In-Hole task. In this work, 2 LfD frameworks are developed and experimented with a number of 20 trials for the Peg-In-Hole:

• Offine LfD – First, the user record the demonstrations (3 demonstrations), learned trajectory is generated, and lastly, the 20 trials are performed using that learned trajectory.
• Online LfD – During any trial of the 20 trials, the user can take over the control (SHARED) and correct or adjust the trajectory which is recorded as a demonstration for the LfD framework. Hence, the learned trajectory is continuously enhanced throughout the 20 trials of the Peg-In-Hole task.

Please refer to the Video for more details