practical control system . In the literature, adaptive con- trol methods incorporating intelligent tools have been widely employed in control systems –. Different from model- basedcontrol approaches, these adaptive control methods could deal with the unknown system dynamics by automat- ically updating laws. Although model-basedcontrol usually shows better control performance, it is based on an accurate model . In practice, an accurate model is hardly to be obtained owing to the complexity of mechanical robot system. In addition, the payload on the manipulator could also make the modelling of the robotic arm more difficult . In , the RBFNN was employed to deal with uncertainties of both robot system and the object, and a prescribed tracking performance was guaranteed. In , NNs were integrated in control design to cope with the uncertain dynamics. In , the adaptive fuzzy control method was used to suppress the uncertain dynamics of the system. Although NNs or fuzzy logic can solve the problem of system uncertainties, these intelligent tools inevitably increase the burden of control system. In order to have good approximation performance, the number of neurons and fuzzy rules will increase, which in turn will bring heavy computation burden to the system and affect the control effect. Moreover, when the learning is insufficient, the inaccurate approximation will influence the control performance. However, among the literatures about adaptive NN control, few studies have concerned with the the way of solving the problems of the insufficient learning condition and computational burden.
Abstract— Touch perception is an important sense to model in humanoid robots to interact physically and socially with humans. We present a neural controller that can adapt the compliance of the robot arm in four directions using as input the tactile information from an artificial skin and as output the estimated torque for admittancecontrol-loop reference. This adaption is done in a self-organized fashion with a neural system that learns first the topology of the tactile map when we touch it and associates a torque vector to move the arm in the corresponding direction. The artificial skin is based on a large area piezoresistive tactile device (ungridded) that changes its electrical properties in the presence of the contact. Our results show the self-calibration of a robotic arm (2 degrees of freedom) controlled in the four directions and derived combination vectors, by the soft touch on all the tactile surface, even when the torque is not detectable (force applied near the joint). The neural system associates each tactile receptive field with one direction and the correct force. We show that the tactile-motor learning gives better interactive experiments than the admittancecontrol of the robotic arm only. Our method can be used in the future for humanoid adaptive interaction with a human partner.
appears to be feasible and robust . Another approach is admittancecontrol, which was introduced by Mason . In a generalized admittancecontrol system, with the measurements of environment force and a desired admittance model, a virtual desired trajectory is obtained and tracked. Then, the compliant behavior is realized by trajectory adaptation. Traditional con- trol method of a robot manipulator is model-basedcontrol, which usually has a good control performance . However, this method heavily depends on the accuracy of a robot model which cannot be guaranteed in many cases. Therefore, adaptive control methods have been widely studied and applied to prac- tical systems . These methods can approximate uncertainties of a system by using tools such as neural network (NN), wavelet network, and fuzzy logic system, etc –. Another key element in admittancecontrol system is the force sensor. Force sensors are regarded as a media for communication between a robot and environment. However, force sensors equipped on the manipulators may cause inconvenience and are usually costly. Due to these reasons, sensorlesscontrol schemes have been received great attention. There are two main methods for estimating the external force: disturbance observer approach and force observer approach based on knowledge of motor torques. In , the disturbance observer approach with knowledge of joint angle has been analyzed. In , a force observer for collision detection based on the generalized momentum has been introduced. In , an collision detection method is first developed for rigid robot arms and other robots with elastic joints.
In addition, the control planning developed for variable speed drives working on PMSM are based of the current control on space vector in a rotor frame. This method requires the knowledge of the rotor shaft position for coordinate transformations and information about the speed. The applications of PMSM drive, the rotor position signal is obtained from a mechanical sensor that will reduce the reliability of the system and added the cost of the drives . Therefore, a strong desire arises in the alternative of PMSM sensorlesscontrol, where the estimators are employed as transducers software or electronic commutation to provide the feedback data required by the control system.
(TSMC) is proposed [12-15]. Nevertheless, still with this improved sliding mode control, one can face two disadvantages. The first is the existence of a singularity point problem. The second is the requirement for any compensation that comes from any drift when the ideal no slip assumption does not hold true. The non-singular terminal sliding mode control proposed in [16-22] solves the first issue. The second issue is raised when we need to include slip into the dynamic of the system. This relaxes the assumption of the pure rolling and approach the motion of the WMR to the pure reality. In the literature, few studies, related to this topic, have been investigated. One of the earliest paper that considered wheel slip as an important aspect can be found in , where a model has been derived considering the adhesion coefficient between the wheels of the robot and the surface as a function of the wheel slip. Many other works considered this issue unavoidable and arrived in publishing important results [24-27].
Permanent magnet synchronous motor don’t need exiting current, operating efficiency and power destiny are both very high, however its high-performance control needs accurate rotor position and rate signal to achieve flux-oriented control. In motion control system of traditional permanent magnet synchronous motor, usually we use photoelectric coding disk or resolver to sampling rotor position and rotor rate. However, these sensor raises the cost of system control, and reduced the reliability of the system. Therefore, it becomes more and more popular to call off these devices to improve the reliability of the system.
In  an ANFIS-Controller was developed for the navigation of a mobile service robot with two differential wheels. It was trained to recognize and follow a line on the floor to its goal position. To realize this the desired motor speed values were predefined in a teacher vector. The fuzzy controller was then adapted in premise and consequence parameters of the rule base via supervised learning and could eventually lead the robot along the line. An approach for an obstacle avoidance controller is given in , where a two-wheeled robot should be able to move in a fixed area without colliding with the objects lying within. The parameters of the membership functions and the consequence part were adpated with supervised learning. The controller was tested in simulation and also experimentally. The authors in  focused on the adaption of the parameters in the membership functions. They predefined all parameters and set a perfect route to a goal position without colliding with an obstacle. This path was then used for adapting position and slope of the membership functions to "smooth the trajectory generated by the fuzzy logic model". The controller was employed sucessfully in a two wheeled, differential driven robot. To decrase the number of rules in one fuzzy controller the navigation algorithm in  is split up in a hierarchial system. Within this, different behaviours like e.g. "goto-target" or "turn-corner" are realized with respective neuro-fuzzy controllers. By doing this they could avoid having one big rule base to include all navigation orders.
When opening doors, robots mainly encounter two issues. The first one is how to recognize and locate the door and the handle in real time precisely in unknown envi- ronments. In order to recognize and locate the handle, vision systems such as laser scanners, cameras and infrared sensors are always used. A few related works realize the recognition of various door handles of unknown geome- tries. Moreno, et al.  investigated different handle types and applied a morphological filter adapted to the charac- teristic shape of different handles to realize the recognition. Klingbeil, et al.  used a computer vision and supervised learning to identify 3D key locations on any handle, thus choosing a manipulation strategy. Ignakov, et al.  extracted the 3D point cloud of any unknown handle by using the optical flow calculated from images taken with a single CCD camera. Most other methods assume the geometry of the handle is already known and the vision systems are just used to locate. Adiwahono, et al.  used a Microsoft Kinect sensor and a 2D laser scanner to esti- mate the handle position, thus planning the trajectory to open the door. Petrovskaya, et al.  presented a unified, real-time algorithm that simultaneously modeled the posi- tion of the robot within the environment, as well as the door and the handle. Kobayashi, et al.  applied an IP camera and IR distance sensors to calculate the position of the handle, which could be cylindrical with its diameter 48 mm to 56 mm or lever type. However, vision systems are frequently subject to calibration errors, occlusions and sight ranges, making it inevitable for scholars to apply force sensing to additionally double-confirm the contact position with the handle [21–23]. In fact, it is completely competent for robots to use only force sensing to detect the positional relationship with the door and the handle by touching at different positions and different directions, just like humans acting in the darkness. To simplify the system and supplement relevant study, it is essential to develop a new door-opening method based on only force sensing. If the robot is far away from the door in an undiscovered room, vision systems , human-computer interaction or some other methods may be applied to help the robot distinguish the door from the wall and navigate the robot to the door, but not involved in measuring the positional relationship.
time-complexities, without sacrificing functionality. To add nodes we developed a mitosis algorithm which causes a parent node to divide into two daughter nodes. A compactness test determines if a node should be divided (mitosis) to better model what might actually be multiple clusters, or if a node should be pruned. The trigger mechanism for mitosis is based on this compactness of a parent node. In general, our compactness test measures how well a node’s model matches the emperical data. If a node has a learned a single cluster, it will not need to undergo midosis. On the other hand, if a node has tried to learn multiple nodes, this will be reflected by a low compactness measurement and, hence, induce mitosis. Two commonly used tests that we considered include: 1) the Kullback-Leibler (KL) test ; and 2) the Kolmogorov-Smirnov (KS) test . We found that both tests had very positive, yet unique properties. However, we desired more control over the sensitivity than that offered by either test by itself. Such control is needed to yield good discrimination between compact and non-compact nodes. Hence, we modified the KS-test by giving it some properities inherent to the KL-test. The resulting new KS-test allows the user to control its sensitivity by applying smoothing and/or bandpass filtering, as well as, averaging and/or likelihood weighting (raised to user-defined exponential powers). We also developed a half- space symmetry (HSS) test which can be used to detect “V”-shaped clusters. Since the compactness tests are based on distances between a data point and a node trying to model it, V-shaped clusters can result in an inappropriate compact measure. Thus, the HSS condition looks for asymmetries about node axes to detect V-shaped clusters and induce mitosis.
In small problems with few states and short prediction horizon, sequential approach is probably more effective (Peña, 2002). Generally, in big problems simultaneous approach is more robust, because it is less probable that it fails. In sequential approach the addition of restrictions in states or outputs is more complicated. In addition to model restrictions, control actions restrictions, steady state restrictions, etc, can be added.
determine the optimal duty ratio to meet the torque and flux command [22-24]. In this paper CDTC first improved using Fuzzy logic based switching controller called Fuzzy DTC (FDTC). As Explained in the [25-27] speed sensor elimination was important in the industry. The Sensorlesscontrol results in lower cost, reduced hardware complexity. The current based Model Reference adaptive system (MRAS) proposed by Jakub Vonkomer et.al.  For a wide speed range of operation is employed in this paper for estimation of motor speed. Thus many researchers improved CDTC using Fuzzy logic or Duty ratio controller. In this paper the combination of both FDTC with ANN and Fuzzy based duty ratio controller for Sensorless direct torque control of induction motor is proposed. So based on literature the CDTC is first improved by a fuzzy logic controller (FLC) which employs 180 fuzzy rule base. The Switching vector of the Voltage Source Inverter (VSI) is chosen based on Fuzzy logic, which is tuned according range of flux and torque errors. The concept of duty ratio is introduced to reduce torque and flux ripple. The duty ratio is developed based on two AI techniques Fuzzy logic and Neural Networks. The speed and torque of the induction motor are estimated based current based adaptive estimator scheme as suggested in . The performance of FDTC with Fuzzy duty ratio (FDR) and Neural Networks based duty ratio (NNDR) is assessed based on simulation results using Matlab/Simulink Software. Simulation results indicate that there is considerable reduction in torque and flux ripples using FDTC with Fuzzy or NN based duty ratio controllers. In both FDTC with Fuzzy DRC and NN DRC the speed of induction motor tracks the reference speed without any peak over shoots as in the case of CDTC. The stator current drawn by proposing schemes contains fewer harmonic and more sinusoidal waveforms compared to CDTC.
planning or visibility graphs. On the other hand, we have the option of using an intelligent control system based on artificial intelligence techniques such as neural networks, fuzzy control, etc. These systems are able to solve many control problems due to its ability to emulate some human characteristics. In , a collision-free path between the source and the destination is constructed based on neural networks for mobile robot navigation in partially structured environments. The proposed scheme uses two neural networks for the task. The neural network multilayer perceptron (MLP) using back- propagation. The proposed scheme is carried out in real time implemented on an Intel Pentium 350 MHz processor; the robot is able to avoid all obstacles to reach its goal from the initial starting point.
The control and the modeling of cable-driven parallel robots are complex due to the flexible nature of the cables. In , the authors proposed controllers for the cable suspended robot using Lyapunov theory and feedback linearization. In this work, there is no measure of Cartesian position. The pose (position and orientation) of the mobile platform can be estimated through the forward kinematics but the latter is generally difficult to solve and the measures of cable lengths by means of active joint variables may be inaccurate due notably to cable elongations. Another solution is to use ad-
Either method conveys adequate information about driving the motor phases based on Hall Effect sensor states. The relationship between the Hall Effect sensors themselves is always consistent. In other words the Hall Effect sensor sequence seen in can be found in all motors with 120- degree Hall Effect sensors when the motor rotates. However, the direction of rotation, CW or CCW, necessary to produce this relationship can vary across different motors. Very often the binary state of the three Hall Effect sensors will be combined to create a 3-bit binary word. The mapping between the Hall states and the three-bit word. Below the binary word representation in is tables that represent the states of the MOSFETs of the half-bridges.
information of the scene into the image space. The camera mounted on the hand of the robotic manipulator reduces the possibility of occlusion and also gives the opportunity to use a fastened-on end- effector. In this paper the method of the force and vision sensors data acquisition and data analysis is described. The extracted data are integrated into a proposed task frame. The proposed method was validated by different tests that were carried out on a testbed described in the section called Experimental system. In the Conclusion the findings and some challenging future improvements of our system are discussed.
Abstarct—This paper builds up the brushless dc (BLDC) sensorlesscontrol framework .The sensorless systems that depend on a hysteresis comparator and a potential start-up technique with a high beginning torque are proposed. The hysteresis comparator is utilized to make up for the stage deferral of the back EMFs because of a low-pass channel (LPF) and likewise keep numerous yield moves from commotion or swell in the terminal voltages. The rotor position is adjusted at stop for greatest beginning torque without an extra sensor and any data of engine parameters. Likewise, the stator current can be effectively balanced by adjusting the beat width of the exchanging gadgets amid arrangement. A few examinations are actualized on a solitary chip DSP controller to exhibit the attainability of the proposed sensorless and start-up strategies.
The main drawback of the FOC is that the shaft speed (or position) feedback is required. This presents a huge problem for low cost systems in which motor mechanical sensors are not available. This has led to sensorlesscontrol of AC machines, a field of research during the past decade [3-13]. Sensorlesscontrol of induction motors has faced two kinds of methods: the one which uses the dynamic model of the induction machine based on the fundamental spatial harmonic of the magnetomotive force (mmf) [4–11], and the other based on the saliencies of the machine   which are based on high-frequency signal injection. Among the first, the main one is the open-loop speed estimators , MRAS (model reference adaptive system) speed observers  , full-order observers  , and reduced order speed observers [9-11]. Most of these methods are applied with vector control by rotor field orientation and are based on the different complex models which require large computation time. The full-order observer gives very
washing machine. Particularly, the automobile industry will increasingly use electric systems as a replacement for hydraulic or pneumatic systems because of their limited maneuverability in applications of the engine management, such as for the electric power steering in which brushless drives are currently implemented. Thus, the brushless permanent magnet motors, PMSMs, are taken into account in this thesis. However, a major disadvantage of these machines is the requirement of a sensor system to detect the rotor position. This sensor requiring extra space and cabling will lead to additional costs. A diagram of a PMSM closed loop controlbased on the field oriented control (FOC) is depicted in Fig. 1.1. It is shown that the speed and position can be usually found by traditional measurements, such as resolvers and absolute encoders.
The results show that the gesture control program can capture gesture images through the Kinect camera, recognize the five kinds of common static gestures input from the camera, and then convert them into corresponding control commands to control TurtleBot2 to move forward, backward, turn left, turn right and stop.