兩足行走機(jī)器人行走控制系統(tǒng)設(shè)計

兩足行走機(jī)器人行走控制系統(tǒng)設(shè)計,兩足行走機(jī)器人行走控制系統(tǒng)設(shè)計,行走,機(jī)器人,控制系統(tǒng),設(shè)計 Comparing Control Strategies for Autonomous Line-Tracking Robots Luis Almeida, Alexandre Mota, Pedro Fonseca Ida, alex,pJ tia.pt Departamento de Electronica e Telecomunica5es Universidade de Aveiro, P-3 8 10 Aveiro, Portugal tel.: +351-34-370 859 fax: +351-34-381 128 Abstract Autonomous mobile robotics is a very exciting area for students particularly for those who attend courses on electronics. The authors have been involved in several activities in this area together with students of the University of Aveiro. In particular, one of stich activities is to build robots to follow a line drawn on the jloor. In order to do this efficiently a simulator has been implemented and used to test the influence of different control approaches. This article presents a brief description of the MATLAB- -based robot model and line-tracking simulator. It then compares several drfferent control approaches in terms of restilting integral absolute error (IAE) and integral squared error (ISE), easiness of tuning and complexify of the respective code. The compared approaches are: proportional, proportional-derivative, proportional- -integral-derivative, f.zzy, table-based fuzzy, self-organi- sing fiizzy and neural-net inverse model based. 1 Introduction Building autonomous robots is an interdisciplinary activity and thus has a great pedagogical value. With this fact in mind the authors have been supporting student teams from the University pf Aveiro to participate in an annual event that takes places in France where, basically, autonomous mobile robots have to follow a line, amongst other tasks To better understand the behaviour of the line-tracking robot and to show students how different scientific concepts from physics, geometry, electronics, instrumentation and control are integrated when building such a robot, the authors have developed an analytical 0-7803-4484-7198 I $10.00 0 1998 IEEE 542 model of the line-following robot l. The model takes into account several real-world constraints and allows to predict the movement of the robot based on the electrical voltages applied to the motors. Also in l the authors have described the geomet? of the line-tracking process which was used to build a simulator. This allowed to determine the path of the robot as well as the relationship between this path and the reference path that the robot is to follow. The simulator is a valuable tool to compare different control approaches, as well as different sensor layouts, prior to the robot construction. This allows for better decisions concerning the physical properties of the robot before actually building it. In the next section this article presents a brief description of the robot model and of the simulator: In section 3 several control approaches are compared, namely propor- tional, proportional-derivative, proportional-integral- derivative, fuzzy, table-based fuzzy, self organising fuzzy and neural-net inverse model based. Conclusions are drawn in section 4 which also includes some comments concerning on-going work. 2 Simulating the Robot 2.1 The Robot Model The robots which have been built by the students in the activities mentioned before are normally simple (fig 1) Motion is achieved by using two independent DC electric motors that drive one wheel each. Differential drile i5 used to steer the robot. One or two extra caster wheels are AMC 98 - COIMBRA used to keep the robot horizontally. The deviation of the robot from the reference path is measured by a set of sensors placed ahead of the robot which are, normally, infrared-light detectors. Typically, closed-loop control of the wheels velocity has not been done. The velocity of each wheel is controlled indirectly by applying voltages to the motors. This option may decrease the performance of the tracking algorithm but simplifies the final tuning. Remember that the use of closed-loop wheel speed control would require the tuning of two extra independent loops. Figure 1. The basic robot These characteristics have been used to derive a model for the line-tracking robot (fig. 2). To improve its accuracy the model takes into account inertia (mass (M) and moment of inertia (a), friction coefficients (for translational (By) and rotational (B,) movements), electric motors parameters (the resistance (R) and the motor constant (Km), additive noise (in the sensor readings) and physical limitations of the robot such as the length of the line-sensors (5) and the maximum voltage that can be applied to the motors (V). The model is described in l and allows to calculate both linear (v) and angular (0) velocities of the robot based on the voltages applied to the motors (Vow - average, and V, differential). 2.2 The Line-Tracking Simulator The robot model referred to above, was complemented with a geometric analysis of the line-tracking problem. This problem falls within the general path-tracking problem which has been treated in the literature, e.g. 2. In particular, the simulator presented in this article follows a reactive approach to track an unknown line as opposed to the planning approach of tracking a path previously planned and thus, known in advance. In l a geometric analysis is also shown that allows to calculate the next deviation from the line (e) based on the present deviation, wheels velocities and angular position of the robot relative to the line. The robot is used as referential. However, in order to better define the reference trajectory and to visualise the robot trajectoq, another model was built in whch the robot position was referred to an absolute referential. In this geometric model, the next deviation from the line (e) is calculated based on the robot absolute position and the wheels velocities. Knowing the robot position (xo,yo,ar it is possible to calculate the intersection of the sensor array with the line (xey,) which then allows to calculate the deviation e (fig. 3). The resulting dependency of e relative to the position of the robot is non-linear. The velocities are used to calculate the robot displace- ment (dZ,.,da,) during an infinitesimal time interval (200. was found that best results were obtained with K,=-tOO and Kp380. Figure 8 shows the deviation obtained along the reference path with these values. The absolute maximum deviation is 23 mm and the IAE is 6.6. A note worth referring is the fact that the controller is capable of converging to zero deviation over straight segments but in curves with constant radius, the deviation converges to a non-zero value. Since the angle of the reference path is constant in straight segments (step input) and increases constantly in curveswith fixed radius (ramp input) the line-tracking robot can be considered as a type 1 system. The same happens with the proportional controller. 3.3 Proportional-Integral-Derivative This type of controller, known as PID, results from the previous one by adding an integral term to the actuating signal. This allows to bring the deviation to zero over any part of the line, either straight or curve. The deblation . . . -0.0, . 0 200 400 600 800 1000 -0.02 Time (Sampling IntONalS) Figure 8. Using a PD controller with Kp=400 and K380. 545 can be kept very small when the right parameters are used although it always increases in the beginning and ending of any curve. The controller output is vdl,=Kpe+Kdce+KII(e). With Kp=200, Kp200 and K,=lOO it was possible to decrease the IAE (7.8) and the maximum absolute error (25 mm). No attempt was done to find the best 3 values. Anyway the results are better than with the non-optimised PD controller. Although controllers of this type normally achieve a good performance, the tuning of the 3 constants is very difficult. The use of non-optimal constants may cause a considerable degradation in performance. 3.4 Fuzzy Logic Approach The fuzzy logic approach can be an alternative to the previous strategies. Although it is more complex than either P, PD or PID approaches, it is still relatively easy to implement since it is based on intuitive rules explicitly given by the programmer 4. In this case a fuzzy incremental controller with normalised universes of discourse and gaussian membership functions is used 51. The controller inputs are the trajectory error (e) and its derivative (ce). The controller output is the differential voltage VdiP The fuzzy control surface can be depicted on figure 9. Note the non- -linear surface and the gradient near the center. Two approaches were tried with thls type of controller: rule-based and table-based algorithm. The first one uses functions from the Mathworks Fuzzy Logic Toolbox. The second is only a 2D Look-Up Table. The results are identical in terms of IAE, ISE and maximum absolute error. Some better results were obtained adding a linear integral term to the fuzzy algorithm (see table 1 for -1 -1 Figure 9. Fuzzy Control Surface. details). However the results were a little bit more modest than the ones obtained with the optimised PD controller. Trying to improve these results lead to the use of a self- organising fuzzy controller. 3.5 Self-Organising Fuzzy Approach The self-organising fuzzy controller (SOC) uses some kind of performance measure to update the rule base. The most common approach has a hierarchical structure in which the lower level is a table-based controller. The higher level monitors the error and the change in error and modifies the table, when necessary, through a modifier algorithm 6. The performance measurement is carried out using expression 2. P is the performance measure or the penalty, that is added to the control table, e is the error and ce is the change in error. Kce is a time constant and Gp is the learning rate factor. Starting with a table similar to the one used on the table-based controller it is possible, after 10 training sessions (of one full referece path each), to improve the overall performance up to the one obtained with the optimised PD controller. Figure 10 shows the IAE evolution along the 10 training sessions.Note that the training occurs on-line while the robot is actually moving along the line. As well as with the simple fuzzy approaches, the addition of an integral action to the SOC allows to achieve even better results as can be seen in table 1. 3.6 Neural Networks Approach Knowing that the robot model predicts non-linear, IAE Evoldon Trapctonms Figure 10. Self-Organising Controller IAE evolution 546 stable, dynamic behaviour lead to the idea of using some kind of neural-network approach in order to implement a direct inverse control algorithm. The inverse model was identified by the use of a 2 layer feedforward network with 4 inputs, 8 hidden non-linear neurons and a linear output neuron. The network was trained off-line with the Levenberg-Marquardt method 7 and, after 5000 epochs, it was possible to get a good inverse model. With the obtained network a direct inverse control scheme was implemented SI. The results obtained this way are the best ones among the compared control strategies as can be seen in table 1. 4 Conclusions Table 1 presents the results obtained with each con- trolling approach. Two main sorts of controllers were used, those capable of learning (SOC, SOC+I and NN) and the remaining ones. From these latter ones it is possible to see that the use of fuzzy controllers does not bring along an immediate benefit. A simple hand- -tuned PD controller performs better. When an integral component is added to the fuzzy controllers, their performance is improved up to the one of the PD controller. However, the V, parameter is still superior in the PD approach. Notice that a difference of 0.03ms- yields a difference of 10s after 30m. Optimally tuning a PD controller is either very difficult (highly time consuming) or even impossible when there is no analytical model of the robot and the tuning has to be done with the real robot. The PID approach is also difficult to tune and, in many cases, the resulting performance may even be worse than for the PD. The tuning of the fuzzy controllers is easier to achieve since it is embedded in the intuitive rules explicitly given by the programmer (rule-based approach) or in the table Table 1. Comparing different control strategies to track an unknown line. Controllzr IAE ISE Vmean Emax/ %IAE/P P 15.1 0.33 0.30 0.049 1 PD PD(optimized) PID FUZZY FUZZY+I FUZZYTABLE FUZZYTABLE+I SOC (10 epochs) SOC+I (15 epochs) NN (SO00 epochs) 9.5 0.14 0.32 0.033 37 6.6 0.07 0.30 0.023 56 7.8 0.09. 0.31 0.025 48 10.8 0.21 0.31 0.051 28 9.5 0.14 0.29 0.037 37 10.7 0.21 0.31 0.053 29 9.5 0.13 0.29 0.034 37 6.6 0.07 0.31 0.024 56 6.2 0.06 0.30 0.021 59 5.2 0.04 0.31 0.021 66 built from such rules (table-based approach). Concerning the controllers capable of learning, the SOC presents a good strategy to improve performance with a relatively low computational cost. Besides, its on-line learning capability as well as its speed of learning make it very atractive. The neural-net-based approach is very powerful but very time consuming (off-line training) and requiring large computational resources (floating-point calculations). Thus, it is not well suited to be used with low processing power microcontrollers. The authors are currently working, together with students, in the construction of a new line-tracking robot (RUA v3.0). Several controllers will be used and experimental data will be collected in order to verify the results presented in this article and obtained with the simulator described in section 2. Particular attention will be given to the use of the SOC controller. References I Almeida L., 1997. Modelizao de Pequenos Robots Autbnomos, Revista do Dep. de Electrbnica e Telecomunicag6es da Universidade de Aveiro, Aveiro, Portugal, Vol. 2, no. 1. 2 Sampei M. et al., 1993. Arbitrary Path-Tracking Control of Articulated Vehicles Using Non-Linear Control Theory, IEEE Transactions on Control Systems Technology, Vol. 3. 3 Vieira J. A. et al, 1997. Moliceiro - Um Robot que Segue Uma Linha, Revista do Dep. de Electrbnica e Telecomunicagi3es da Universidade de Aveiro, Aveiro, Portugal, Vol. 1, no. 7. 4 Rausis K. et al., 1997. Obstacle Avoidance Control of a Mobile Robot, Proceedings of the IFAC - SICICA 9 7 Symposium, Annecy, France. 5 Jan Jantzen, Fuzzy Control, Lecture notes in On-Line Process Control, Technical University of Denmark, 199 1. 6 Jan Jantzen, Fuzzy Control Course On The Internet, Technical University of Denmark, http:/wwv iau. dtu. dk/-jj/learn/index. html. 7 Nsgaard, 1997a. Neural Network Based System Identijcation Toolbox, Tech.Rept. 97-E-85 1, Dep. Automation, Technical University of Denmark. SI Nsgaard, 1997b. Neirral Network Based Control System Design Toolkit, Tech.Rept. 96-E-830, Dep. Automation, Technical University of Denmark. 547