A robotic assembly task is usually implemented as a sequence of simple motions, and the transitions between the motions are made when some events occur. These events can usually be detected with thresholds on some signal, but faster response is possible by detecting the transient on that signal. This paper considers the problem of detecting these transients. A force-controlled assembly task is used as an experimental case, and transients in measured force/torque data are considered. A systematic approach to train machine-learning based classifiers is presented. The classifiers are further implemented in the assembly task, resulting in a 15%reduction of the total assembly time.

For many robot operations force control is needed, but force sensors may be expensive and add mass to the system. An alternative is to use the motor torques, though friction causes large disturbances. The Coulomb friction can be quite well known when a joint is moving, but has much larger uncertainties for velocities close to zero. This paper presents a method for force estimation that accounts for the velocity-dependent uncertainty of the Coulomb friction and combines data from several joints to produce accurate estimates. The estimate is calculated by solving a convex optimization problem in real time. The proposed method was experimentally evaluated on a force-controlled dual-arm assembly operation and validated with data from a force sensor. The estimates were shown to improve with the number of joints used, and the method can even exploit data from an arm that is controlled not to move.

Industrial robots are usually position controlled, which requires high accuracy of the robot and the workcell. Some tasks, such as assembly, are difficult to achieve by only using position sensing. This work presents a framework for robotic assembly, where a standard position-based robot program is integrated with an external controller performing force-controlled skills. The framework is used to assemble emergency stop buttons that were tailored to be assembled by humans.

The traditional way of controlling an industrial robot is to program it to follow desired trajectories. This approach is sufficient as long as the accuracy of the robot and the calibration of the workcell is good enough. In robotic assembly these conditions are usually not fulfilled, because of uncertainties, e.g., variability in involved parts and objects not gripped accurately. Using force control is one way to handle these difficulties. This paper presents a method of doing force control without a force sensor. The method is based on detuning of the low-level joint control loops, and the force is estimated from the control error. It is experimentally verified in a small part assembly task with a kinematically redundant robotic manipulator.

In absence of covariance data, Kalman filters are usually initialized by guessing the initial state. Making the variance of the initial state estimate large makes sure that the estimate converges quickly and that the influence of the initial guess soon will be negligible. If, however, only very few measurements are available during the estimation process and an estimate is wanted as soon as possible, this might not be enough. This paper presents a method to initialize the Kalman filter without any knowledge about the distribution of the initial state and without making any guesses.

Modern industrial robots are fast and have very good repetitional accuracy, which have made them indispensable in many manufacturing applications. However, they are usually programmed to follow desired trajectories and only get feedback from position sensors. This works fine as long as the environment is very well structured, but does not give good robustness to objects not being positioned or gripped accurately. A solution is to use additional sensing, such as force sensors and vision. How to combine the data from the different sensors and use it in a good way to control the robot is still an area of research. This paper describes an assembly scenario where a switch should be snapped into place in a box. Force sensing is used to resolve the uncertain position of the parts and detect the snap at the end of the operation. During the assembly an uncertain distance is estimated to improve the performance. By performing the assembly several times, learning is used to generate feed-forward data, which is used to speed up the assembly.

The use of industrial robots in the aircraft industry has been hampered by a combination of poor accuracy of the robots and poor calibration of the workcell, and also manufacturing variability in composite parts. A way to handle these difficulties is using force control. An experimental case where a semi-compliant rib is aligned to multiple surfaces is used as an example to show this. The constraint-based task specification framework is used for the modelling and control, and the search and alignment sequence required for the assembly is modeled with a state machine. An implementation on an industrial robot system is presented and experimental data is evaluated. The described approach is easy to apply to other fields and more complicated assembly operations as well.

To be able to determine the position of a static object in 3D space by means of computer vision, it has to be seen by cameras from at least two different view points. The same applies for measuring the position of a moving object based on images captured at one single time instant. However, if the cameras are not synchronized in time, or if a moving object is not visible in all images, one can not rely on using matching pictures for making accurate position estimates of dynamical objects. This paper presents a strategy to track an object with known dynamical model, using a series of images where no pair has to be captured simultaneously. It even allows tracking of a point object in 3D space using a single static camera.