Master thesis : Virtual Mechanism Control for Compliant Peg-in-Hole Assembly on an Industrial Manipulator

Abstract

Robotic assembly in industrial environments remains a challenge when parts come into unavoidable contact, as conventional position-controlled manipulators push rigidly against surfaces, risking damage to the assembled parts. This thesis investigates the use of torque control and passive compliance for robotic peg-in-hole assembly, motivated by the industrial needs of Cilyx, a company specialising in production line automation. The Virtual Mechanism Control (VMC) framework is adopted as the control paradigm. Building on the principles of passivity-based impedance control, the VMC approach introduces virtual mechanical elements, namely springs and dampers, interconnected in cartesian space to generate compliant and modular control behaviours. A controller consisting of two virtual spring-damper pairs, one governing translation and one governing orientation, is designed and validated on a Universal Robots UR5e manipulator. The software architecture is built on ROS2 and the ur_driver, providing a modular and transferable control stack that is decoupled from the robot-specific communication protocol, in direct alignment with the standardisation objectives of Cilyx. The controller is first validated in simulation, where the influence of the stiffness and damping parameters on convergence speed and force levels is characterised. The transfer to the real robot immediately exposes a significant sim-to-real gap: joint friction forces the controller to operate at high stiffness to achieve acceptable positioning accuracy, which in turn reduces the passive compliance that torque control is designed to provide. This fundamental tension between accuracy and compliance, caused entirely by friction, is the defining constraint of the experimental results.

Three experiments are conducted on the real robot. In the first, with perfect knowledge of the hole position, a success rate of 100% is achieved across all six peg geometries and both clearance levels tested, confirming the repeatability of the controller. In the second, a deliberate positional offset is introduced without any search strategy, relying exclusively on the passive compliance of the VMC controller to absorb the error.

This experiment fails systematically, demonstrating that the compliance deficit induced by joint friction is insufficient to recover from millimetre-scale positional errors without an explicit search mechanism. In the third experiment, an Archimedean spiral search phase is added to the trajectory, and success rates of 91.6%, 78.3%, and 58.3% are obtained for offsets of 2, 4, and 8 mm respectively, without any force or torque sensing.

These results demonstrate that the proposed controller, which combines a passivity-based VMC layer, a waypoint-based trajectory strategy, and a spiral search phase, constitutes a viable and transferable proof of concept for compliant peg-in-hole assembly on an industrial manipulator. The work identifies friction compensation as the most impactful direction for future improvement, and concrete perspectives are outlined including the extension to position-controlled robots and the integration of force feedback.