On Robotic Assembly using Contact Force Control and Estimation

Abstract: Force sensing provides robots the capability to accomplish tasks where physical interaction with the environment is required, such as assembly. Small position uncertainties can then be corrected for by sensing the contact forces. This thesis considers the problem of force-controlled assembly, including how tasks can be specified in a simple and intuitive way and how robust task execution under uncertainties can be accomplished. A framework for performing robotic assembly is presented. An assembly tasks is composed of a number of skills, where skills both can be force controlled and be carried out using standard position-based control. The skills using force control are specified as sequences of constrained motions, where transitions between the motions are triggered by sensor events. These events can either be simple threshold levels, or be more advanced classifiers based on machine learning. A method for explicitly modeling and resolving uncertainties is presented, as well as a method for adaptation of force control parameters based on identification of a contact model. Specification of sensor-based skills usually requires expert knowledge. To make the specification procedure more easy and intuitive, this thesis presents a method where force-controlled skills can be specified on a high level, and where an executable low-level description is generated. Experimental implementations of multiple assembly scenarios are used to validate the methods and to investigate the potential for force-controlled assembly with industrial robots. A force sensor may not always be available. The thesis presents two different methods for performing force estimation, based on the measured joint motor angles and the joint motor torques. Friction in the joints is the major disturbance when doing force estimation. A method to increase the accuracy of force estimation using dithering to decrease the effective friction level is proposed. Lead-through programming, to manually guide the robot, is useful for simple and intuitive robot programming. The thesis presents a method for performing such lead-through programming without any force sensor, based on disabling the low-level joint controllers, only feedforwarding the torque to compensate gravity. Specification and execution of tasks based on external sensing is difficult for non-experts. The methods presented in this thesis all contribute to making it easier and more intuitive to use industrial robots for performing assembly tasks.