Search

Even with the rapid technological advancements, robots are still not the most comfortable machines to work with. Firstly, due to the separation of the robot and human workspace which imposes an additional financial burden. Secondly, due to the significant re-programming cost in case of changing products, especially in Small and Medium-sized Enterprises (SMEs). Therefore, there is a significant need to reduce the programming efforts required to enable robots to perform various tasks while sharing the same space with a human operator. Hence, the robot must be equipped with a cognitive and perceptual capabilities that facilitate human-robot interaction. Humans use their various senses to perform tasks such as vision, smell and taste. One sense that plays a significant role in human activity is 'touch' or 'force'. For example, holding a cup of tea, or making fine adjustments while inserting a key requires haptic information to achieve the task successfully. In all these examples, force and torque data are crucial for the successful completion of the activity. Also, this information implicitly conveys data about contact force, object stiffness, and many others. Hence, a deep understanding of the execution of such events can bridge the gap between humans and robots. This thesis is being directed to equip an industrial robot with the ability to deal with force perceptions and then learn force-based tasks using Learning from Demonstration (LfD). To learn force-based tasks using LfD, it is essential to extract task-relevant features from the force information. Then, knowledge must be extracted and encoded form the task-relevant features. Hence, the captured skills can be reproduced in a new scenario. In this thesis, these elements of LfD were achieved using different approaches based on the demonstrated task. In this thesis, four robotics problems were addressed using LfD framework. The first challenge was to filter out robots' internal forces (irrelevant signals) using data-driven approach. The second robotics challenge was the recognition of the Contact State (CS) during assembly tasks. To tackle this challenge, a symbolic based approach was proposed, in which a force/torque signals; during demonstrated assembly, the task was encoded as a sequence of symbols. The third challenge was to learn a human-robot co-manipulation task based on LfD. In this case, an ensemble machine learning approach was proposed to capture such a skill. The last challenge in this thesis, was to learn an assembly task by demonstration with the presents of parts geometrical variation. Hence, a new learning approach based on Artificial Potential Field (APF) to learn a Peg-in-Hole (PiH) assembly task which includes no-contact and contact phases. To sum up, this thesis focuses on the use of data-driven approaches to learning force based task in an industrial context. Hence, different machine learning approaches were implemented, developed and evaluated in different scenarios. Then, the performance of these approaches was compared with mathematical modelling based approaches.