How to cope with a closed industrial robot control: a practical implementation for a 6-dof articulated robot

Transcription

1 How to cope with a closed industrial robot control: a practical implementation for a 6-dof articulated robot Basilio Bona Tommaso Calvelli Dipartimento di Automatica e Informatica Politecnico di Torino C.so Duca degli Abruzzi Torino, Italy basilio.bona@polito.it April 15, 2003 tommaso.calvelli@polito.it Abstract Robot manufacturers use simple algorithms (such as PID) to control industrial manipulators. The controller architecture and the algorithm parameters are closed to any change from external users. This is often unacceptable in teaching or research environments. In this paper a simple and inexpensive architecture is presented for the COMAU Smart S2 6dof manipulator, aimed at introducing the possibility to integrate external sensors, such as force/torque sensors or CCD cameras for advanced control tasks. The main difficulty is to cope with the temporal limits introduced by the limited sampling rate attainable by such an architecture. Experiments were performed using a force/torque sensor to test the strengths and limits of the proposed architecture. Keywords : control architecture, closed robot, experimental setup. 1 Introduction Industrial robots are usually controlled by simple HW/SW architectures based, for example, on the classical PID position control algorithm. A reason for this choice is to be found in the fact that a position control is usually sufficient to fulfill the point-to-point accuracy requirements, and an increase in the final position accuracy is preferred to a more complex control algorithm, such as inverse dynamics control or force control. 1

2 Moreover, in industrial manipulators, the controller architecture is commonly closed to external modification, so that it will be very difficult for the user to change any parameter and almost impossible to modify the control algorithms. If this limitation is acceptable and often necessary in an industrial environment, it is much less tolerable in an experimental laboratory where teaching and research activities require to be able to modify the control settings. As a consequence of this closeness, if the user wants to change the control algorithm or add other sensors to the existing ones for particular requirements, hardware changes and/or additions in the controller cabinet are required. This work describes a practical test case that shows how it is possible to set up an open architecture without costly hardware modification. It is based on the Smart S2 industrial robot with C3G 901 controller, both produced by COMAU SpA. On request from research groups, COMAU supplies the proprietary controller unit together with a BIT3 computer adapter board, affording the possibility to communicate at different levels (modes) with an external PC. The modes differ for the speed of communication and the level of interaction with the controller. A closed proprietary C library (PCC3Link produced by Tecnospazio SpA) is available to drive the board, and to read/write all the signals exchanged between the controller and the manipulator. This particular setup does not allow to change either the control architecture or the control parameter values (Tecnospazio, 2000). Unfortunately the PC must work under the MS-DOS operating system (and not under the DOS-Prompt in any version of Windows); therefore the possibilities of interaction are limited to the addition of an external signal to the internal reference signal (Tecnospazio, 2000). The advantage of this architecture is the possibility to interact in real-time with the controller; the disavantages derive from the operating system limitations and the necessity to stay within the time limits imposed by the controller. Furthermore, due to safety reasons, only some of the possible modes are available to the user and consequently only a partial interaction with the controller is possible. This fact motivated the authors to find other solutions within the framework of the limited functionality of the Tecnospazio system. Similar results were published along these lines, using the same hardware (with or without full functionality). In (Bassi et al., 2001) an adaptive modification for the existing PID control algorithm is presented. The adaptive set-point is generated in order to reduce the tracking error and it works in parallel with the PID control. In that work the authors aimed to solve the temporal window problem trying to reduce drastically the computation time, using, for example, Assembler code instead of C code. In (Natale and Siciliano, 1998) and (Natale, 1999) and the references cited herein, the interaction with the controller is performed at the joint current level. Again, the problem is the limited temporal window, and it has been solved improving the speed of sensors acquisition and programs. The main contribution of this paper lies in the description and experimental validation of a simple but new architecture for removing the problem of the 2

3 temporal window and the limits of the MS-DOS, without changing the hardware architecture of the controller. 2 Description of the Improved Architecture The robot is a 6-DOF anthropomorphic industrial manipulator, shown in Figure 1, and in its pure industrial version is available only with its native C3G controller. The scheme illustrating the original robot architecture is shown in Figure 2. Figure 1: The robot tip with force/torque sensor. An additional feature offered by Tecnospazio consists on a link between the C3G-Controller and a Personal Computer (PC1) through a BIT3 board, whose function is to exchange data with the controller. It is compulsory to use the MS- DOS operating system, and the read/write exchange is provided by a proprietary PCC3Link C-functions library. Using this configuration, the controller can be used both in normal mode (no data exchange with the computer) or in semiopen mode. It this second mode, an interrupt signal is sent by the controller to the computer each T s = 20 ms; before the next interrupt signal occurs, the computer must supply the controller with the successive joint or cartesian position setpoint. It must be noted that the interaction with the controller is limited only to the position set-point generation. Indeed, with an optional full open version it will be also possible to interact with the controller at the drives current setpoint level, each T s = 1 ms, using the controller only to give power to the motor drives. This last mode is presently not available in our system due to safety reasons. 3

4 Pc 1 C3G Controller Bit 3 ISA Bus Figure 2: The original architecture scheme. As previously said, a common problem encountered in such an architecture is the temporal constraints. For example, if a visual servoing control system is to be implemented, it is necessary to perform the following steps: image acquisition from a camera, feature extraction from the image and use of these features in the control loop. A low-cost camera can acquire an image with a frame rate of about 50 Hz. This approximately equals the controller interrupt rate; so, if we want to generate a position set-point from a vision system and use this set-point for a generic control algorithm there is no time to perform any image data extraction or elaboration, and the present architecture must be changed. Pc 2 Com 1 Serial link Pc 1 Com 1 C3G Controller Bit 3 ISA Bus Bit 3 Figure 3: The proposed general scheme 4

5 Moreover, the MS-DOS operating system, running on PC1, does not allow to integrate a vision system on it: there is no commercial frame-grabber provided with MS-DOS drivers. Figure 3 presents a general scheme for the adopted solution. In this architecture, PC1 is again connected with the controller: it receives the interrupt signal and sends the position set-point through an ISA bus each time it is required. The new added feature consists in a connection with another computer, denoted by PC2, which generates the position set-point and sends it to PC1 via a serial link. While PC1 must create a new set-point and send it to the controller within specified time intervals, PC2 has no time constraints: if a new set-point is computed, it will be sent to PC1, otherwise it will not. PC1 waits for the new set-point until the end of the allotted time interval and if the set-point from PC2 is absent, a new set-point must be created in any other way by PC1 and transmitted to the controller. Different strategies are possible in this last case, i.e.: a) the previous setpoint is supplied, resulting in a movement halt; b) a multirate scheme, where the intermediate high-rate set-points are computed according to a prediction supplied by an observer or a dynamic filter, using the low-rate target set-point; other mixed strategies are always possible. An approach along these lines has been recently published in (Sim et al., 2002). It is important to note that with the proposed architecture, PC2 can use any operating system supporting a vision system. 3 An Experimental Test Case As a practical case, a simple experiment has been performed at the Robotics Laboratory of Politecnico di Torino, to test the architecture functionality and the serial links timing. In this section a description of the hardware used and the links characteristic will be given. The goal of the proposed architecture is to allow the addition of new cartesian sensors, such as force/torque or vision ones, to use them in advanced control laws. Given the necessity to investigate the characteristics and the performances of the proposed set-up, a test was preliminarily conducted using an external force/torque sensor. Such sensor impose a set of constraints that are similar to those of the vision control, at least for the aspects of interconnection and set-point generation we are interested in. The COMAU Smart robot has been equipped with an ATI Gamma SI force/torque sensor (see again Figure 1), mounted on the robot tip, in order to create a force feedback loop. The sensor is linked to PC2 by a serial connection at baud, while the link between PC1 and PC2 is interrupt driven, at the highest possible bit rate for a serial connection ( baud). In our test-case the PC1 is a Pentium III at 500 MHz; the PC2 is a 486 at 66 MHz. The following operations are executed at each sampling period on PC1: 5

6 Force sensor transducer Pc 2 Com 1 Serial link Pc 1 Com 1 C3G Controller Bit 3 ISA Bus Bit 3 Figure 4: The proposed final scheme read the motor positions from the C3G controller; send the positions to PC2 through the serial connection; check whether a new position set-point is available from PC2. If not, PC1 creates a predicted set-point or stops the movement: in both situations the C3G controller receives the new set-point in the correct temporal window. The generation of the set-points on PC2 follows from the particular choice of the desired robot behavior; in our experiments we decided that the manipulator should move in order to counter-react a variable force/torque applied at the endeffector by the operator s hand. If no force/torque is applied, the manipulator shall not move; if it is pushed or pulled or a torque is applied, the manipulator shall move in order to annihilate these external inputs, retracting, advancing and/or rotating according to the sign of the force/torque. In the following Section a brief description of the algorithm used to generate the set-point is presented. 3.1 Set-point generation Let q(k) be the vector of the six actual joint positions at time kt, where T = 20 ms is the sampling period. Let R b and R e be the base and the end-effector reference frames respectively, with R b e the rotation matrix from R b to R e, i.e. R e represented in R b ; let d b e be the translation vector from the origin of R b to the origin of R e. Let p(k) be the vector of the six cartesian parameters (3 position parameters + 3 attitude parameters), using the robot forward kinematic function 6

7 p(k) = F kin (q(k)) we can compute the matrix Te b (k), where ( ) R b Te b e (k) d b e(k) (k) = 0 T 1 is the homogeneous matrix representing the end-effector frame in the base frame. From the force/torque sensor we read both the actual force vector f e (k) and the torque vector τ e (k), expressed in the end-effector frame R e. These vectors are then represented in the base frame R b by: f b (k) = R b e(k) f e (k) τ b (k) = R b e(k) τ e (k) From f b (k) we compute the new position d b e(k +1) of the tool in the base frame: d b e(k + 1) = d b e(k) + K 1 f b (k) (1) where K 1 is a diagonal matrix of gains k 1i, expressed as the inverse of elastic constant, m N 1. The new rotation R b e(k + 1) is computed from τ b (k) as follows: since any rotation can be regarded as a rotation about an axis u, by an angle θ, expressed by (Bona and Indri, 2000) R(u, θ) = I + sin(θ) 1 cos(θ) S(u) + u u 2 S(u) 2 (2) where S(u) is the skew-symmetric matrix function of vector u, we compute separately the new angle θ(k + 1) and rotation axis u(k + 1), from the following expressions: θ(k + 1) = k θ τ e (k) ; (3) τ e (k) u(k + 1) = u(k) + k τ (4) τ e (k) where k θ and k τ are suitable constant gains Next we compute the rotation matrix R e (k + 1) from (2), and we use (3) - (4) to produce the new attitude target R b e(k + 1) = R b e(k)r e (k + 1) Finally we can write the updated homogeneous matrix as ( ) R b Te b e (k + 1) d b e(k + 1) (k + 1) = 0 T 1 (5) and extract from it the updated cartesian parameters p(k + 1). After that, the new joint set-point is computed using the inverse kinematic F 1 kin (p(k + 1)): q(k + 1) = F 1 kin (p(k + 1)) 7

8 and then sent to PC1 using the serial link connection. It would have been possible to use more general upgrade laws than those specified in; for example: A d (z 1 )d b e(k) = B d (z 1 )f b (k) A θ (z 1 )θ(k) = B θ (z 1 ) τ e (k) A u (z 1 )u(k) = B u (z 1 )τ e (k) (6) where z n is the shift operator z n q(k) = q(k n), and A d, A θ, A u, B d, B θ and B u are suitable polynomial matrices in z 1. In the present paper only the simple laws (1), (3) and (4) will be tested. Another possibility for computing the upgraded rotation matrix, instead of considering separately the angle and the axis as in (3) and (4), is to write R b e(k + 1) as R b e(k + 1) = R(dα)R b e(k) (7) where the global error matrix R(dα) is obtained as R(dα) = K α (τ(k)) (8) and K α is a suitable (constant or dynamic) gain matrix. Quaternion characterization, as outlined in (Wen and Kreutz-Delgado, 1991) and (Caccavale and Siciliano, 2001) has been considered as a possible alternative solution to the update of the attitude parameters, but it will not be tested here. 3.2 Experimental results To produce different variable forces and torques on the robot tool-center point, an operator randomly pulls or pushes it: the robot then should move to counteract the forces/torques in order to reduce them to zero. The sampling rate of the internal control loop is 1 ms, while the proposed external controller has a sampling rate of 20 ms. Figure 5 reports the acquired force measurements along the x-axis of the base frame, expressed in the base frame. Figure 6 shows the resulting x-axis cartesian position of the end-effector, moving under the effect of the applied force. The position is obtained computing the forward kinematic function of the measured joint values. A short video, showing the robot behaviour under this control scheme, can be seen at the site /Force/index.htm 4 Discussion and future work The main motivation behind this work is to design an additional set-up in order to overcome in a simple manner, i.e. without problematic re-engineering and 8

9 Measured Force (N) Time (s) Figure 5: The force measured along x-axis [N] x cartesian position (mm) Time (s) Figure 6: The position measured along x-axis [mm]. expensive hardware modification, the constraint imposed to the users by a closed industrial controller. The results obtained are satisfactory and suggest that the next step (i.e. the visual set-point generation) can be performed without much problems, apart those stemming from the feature extraction algorithms. Some problems were nonetheless encountered using the serial link connection; it is well know that this kind of link is affected by errors during the data transmission. Despite the fact that no error correction was performed during experiments, it was observed that during several minutes of transmission only few set-points were lost. The choice to use native Tecnospazio Bit3 adapter and PCC3Link library was, in this phase, compulsory and strongly conditioned the results, due to the intrinsic limitations of MS-DOS operating system. A different solutions to the main problem will be to use a Real-Time Oper- 9

10 ating System (RTOS) on PC1 (our preferences go to RTAI Linux). In this case all the tasks could be implemented on the same PC, without the need of the link with a second PC, as in our test case. The problem with such an architecture will be the necessity to write a low-level RT driver to communicate with the controller through the BIT3 board. 5 Conclusion In this paper a simple architecture for overcoming the limitations of a closed controller in an industrial robot has been illustrated. The main feature of this architecture is the possibility to insert in the control loop additional tasks requiring a computation time longer than the native controller sampling time. Such a task can be, for example, a force control or a vision-in-the-loop control. Apart from additional PC based communication link supplied by Tecnospazio SpA, no other significant hardware modification was necessary. The experiments performed, taking into account force-torque sensor readings, showed the functionality of the architecture, particularly the data transmission between the two PCs and the external slow sensor. Acknowledgments The authors acknowledge the financial support of MIUR under MISTRAL and MATRICS National Research Projects, and ASI under ARS and I/R/137/01/78 Projects. References Bassi, E., F. Benzi, A. Braga and M. Trabatti (2001). Pid regulators for industrial robot. Automazione e Strumentazione 11, (in Italian). Bona, B. and M. Indri (2000). Modelling, Planning and Control of Industrial Robots. Politeko. Torino. (in Italian). Caccavale, F. and B. Siciliano (2001). Quaternion-based kinematic control of redundant spacecraft/manipulator systems control of redundant spacecraft/manipulator systems. In: IEEE International Conference on Robotics and Automation. pp Natale, C. (1999). Six-DOF Interaction Control of Robot Manipulators. PhD thesis. Università di Napoli Federico II. Natale, C. and B. Siciliano (1998). Experiments of visual servoing on an industrial robot. In: IEEE Mediterranean Conference on Control. Vol. 34. Alghero. 10

11 Sim, T.P., G.S. Hong and K.B. Lim (2002). Multirate predictor control scheme for visual servo control. IEE Proc. Control Theory Appl. 149(2), Tecnospazio (2000). Manuale Utente Pcc3Link. Tecnospazio SpA. Milano. in Italian. Wen, J.T.-Y. and K. Kreutz-Delgado (1991). The attitude control problem. IEEE Trans. on Automatic Control 36(10),