The teach and playback mode has limited flexibility  as it is unable to adapt to the many problems that might be encountered in the welding operation, for example, errors in pre-machining and fitting of the workpiece, and in- process thermal distortion leading to change in gap size。 Thus, advanced applications of robotic welding require an automatic control system that can adapt and adjust the welding parameters and motion of the welding robots (Hongyuan et al。 2009)。 Hongyuan et al。 (2009) developed a closed loop control system for robots that used teach and playback based on real-time vision sensing for sensing topside width of the weld pool and seam gap to control weld formation in gas tungsten arc welding with gap variation in multi-pass welding。 In spite of all the above- mentioned drawbacks, online programming is still  the only programming choice for most small to median enter- prises (SMEs)。 Online programming methods using more intuitive human-machine interfaces (HMI) and sensors in- formation have been proposed by several  institutions (Zhang et al。 2006; Sugita et al。 2003)。 The assisted online programming can be categorized into assisted online pro- gramming and sensor-guided online programming。 Al- though  dramatic  progress  has  been  carried  out  to   make

online programming more intuitive, less reliant on oper- ator skill, and more automatic, most of the research outcomes are not commercially available aside from Sugita et al。 2003。

Offline programming Offline programming (OLP) with simulation software allows programming of the welding path and operation sequence from a computer rather than from the robot itself。 3D CAD models of the workpieces, robots, and fixtures used in the cell are required for OLP。 The simulation software matches these 3D CAD models, permitting programming of the robot’s welding trajectory from a computer instead of a teaching pendant in the welding cell as in online programming。 After simulation and testing of the program, the instructions can be exported from the computer to the robot controller via an Ethernet communication network。 Ongoing research sug- gests, however, that the use of sensing technology would make it feasible to completely program the final trajectory only with OLP (Miller Electric Mfg Co。 2013)。 Pan et al。 (2012a) developed an automated offline programming method with software that allows automatic planning and programming (with CAD models as input) for a robotic welding system with high degrees of freedom without any programming effort。 The main advantages of OLP are its reusable code, flexibility for modification, ability to gener- ate complex paths, and reduction in production downtime in the programming phase for setup of a new part。 Never- theless, OLP is mostly used to generate complex robot paths for large production volumes because the time and cost required to generate code for complex robotic systems is similar to if not greater than with online programming (Pan et al。 2012a)。 Currently, for a complex manufacturing process with small to median production volume, very few robotic automation solution are used to replace manual production due to this expensive and time-consuming pro- gramming overhead。 Although OLP has the abovemen- tioned advantages, it is not popular for small to median enterprise (SME) users due to its obvious drawbacks。 It is difficult to economically justify an OLP for smaller product values due to the high cost of the OLP package and pro- gramming overhead required to customize the software for a specific application。 Development of customized software for offline programming is time-consuming and requires high-level programming skills。 Typically, these skills are not available from the process engineers and operators who often perform the robot programming in-process today。 As OLP methods rely accurate modeling of the robot and work cell, additional calibration procedures using extra sensors are in many cases inevitable to meet re- quirements (Pan et al。 2012b)。