s in the Loop: Supporting an Incremental -based Design Process Xiaolin Hu Computer Science Department Georgia State University Atlanta, GA, USA xhu@cs.gsu.edu Abstract This paper presents the results of applying an incremental simulation-based design process to study a robotic convoy system. -in-the-loop simulation, as a major step in this process, allows the system to be measured with combined robot models and real robots. This capability effectively bridges the gap between conventional simulation where models are used and real system experiment where real robots are used. For each step in this incremental process, the simulation/experiment setup is described. The measurement data are then presented and compared. These experiments and results demonstrate the capabilities of robot-in-the-loop simulation and justify the effectiveness of using the incremental simulation-based design process. Keywords: -in-the-loop simulation, Incremental Design Process, ic Convoy, DEVS 1 Introduction Distributed robotic systems usually include a large number of robots that communicate with each other to achieve coordination. Due to the complexity of these systems, verification and evaluation are important in the design process to check if the system under development will fulfill correct behaviors and achieve desired performance. Traditionally, simulation plays important roles from this perspective. However, these roles are typically constrained to the model world on computers. When real system components, such as real robots, are brought into the development, the simulation models quickly become outdated and are hardly reused. Instead, real system experiments are carried out where real robots are tested in a real field. This transition from simulation models to real system components is a necessary step. Unfortunately, it is rarely smooth due to the fact that there exist discrepancies between simulation models and real system components. For example, a robot model used in simulation may not model well (or it is very hard to model well) the mechanic dynamics of a real robot. This difference between simulation models and real system components results in a gap between simulation-based study and real system experiment. Such a gap is significant Narayanaswami Ganapathy Bernard P. Zeigler Arizona Center for Integrative Modeling and University of Arizona Tucson, AZ, USA. {narayang, zeigler}@ece.arizona.edu for large-scale multi-robot systems that operate in challenging environments. To smooth the transition from conventional simulation to real system experiment and to bridge the gap between them, we developed a simulation-based virtual environment that allows combined real robots and virtual robot models to work together [1]. We call this capability of including real robots into simulation robot-in-the-loop (RIL) simulation. RIL simulation brings simulation-based study one step closer to the reality and allows system-wide test to be carried out using combined models and real robots. This is especially useful for large-scale cooperative robotic systems whose complexity and scalability severely limit experimentations in a physical environment using all real robots. For large-scale cooperative robotic systems that include hundreds of robots, RIL simulation makes it possible to conduct system-wide tests and measurements without waiting for all real robots to be available, because the rest of the robots can be provided by the simulationbased virtual environment. The capability of RIL simulation adds into conventional simulation and real system experiment to form an incremental measurement process that includes three steps: conventional simulation, RIL simulation, and real system experiment. As the process proceeds, the system under development evolves from models, to combined models and real robots, and to all real robots. In this paper, we show how this process is applied to a cooperative robotic convoy system. The setup of each step is described and some simulation/experimental results are presented to demonstrate the feasibility and effectiveness of this process. This research is an extension to our previous work on model continuity [] and simulation-based virtual environment [1]. The models and simulation environment that were developed are based on the Discrete Event System Specification (DEVS) modeling and simulation framework [3]. The rest of this paper is organized as follows. Section presents the incremental measurement process. Section 3 describes the robotic convoy example. Section describes simulation/experiment setups and presents the measurement results. Section discusses this research and provides some future research directions.
Incremental Measurement Process An incremental measurement process is formed by integrating RIL simulation that allows combined robot models and real robots to work together. This process includes three steps: conventional simulation, RIL simulation, and real robot experiment. Figure 1 illustrates this process by considering a system with two robots. Model Virtual Virtual Actuator Environment Model (a) Model Virtual Virtual Actuator Model Virtual Virtual Actuator Virtual HIL Actuator Environment Model (b) Actuator Environment Figure 1: An incremental measurement process (c) Actuat The first step is conventional simulation, where all components are models that are simulated on computers. As shown in Figure 1(a), in this step both robot models are equipped with virtual sensors/actuators (sensor/actuator models, which are implemented as abstractactivities in DEVS) to interact with an environment model. Couplings between two robots can be added so they can communicate with each other. This conventional way of simulation has the most flexibility because all components are models thus different configurations can be easily applied to study the system under development. The second step is RIL simulation where one or more real robots are included together with other robot models that are simulated on computers. By replacing robot models with real robots, this step brings simulation-based study closer to the reality and increases the fidelity of simulation results 1. As shown in Figure 1(b), in this step the robot model still use virtual sensors/actuators. However, depending on the study objectives, the real robots may have a combination of virtual sensors/actuators and HIL (from Hardware-In-the-Loop) sensors/actuators. A HIL sensor/actuator, implemented as HILActivity, acts like a real sensor/actuator, but is also coupled to the environment model to synchronize with it. For example, a HIL motor will drive a real robot to move in a real world. Meanwhile, it sends messages to the environment model to update its position in the virtual world. More information about HIL sensor/actuators can be found at [1]. In Figure (b), the real robot uses a virtual sensor and a HIL actuator. Through the virtual sensor, it gets sensory input from the environment model. Using the HIL actuator, it interacts with a real environment (which is not shown in the figure) and also synchronizes with the environment model. In RIL simulation, robot models and the environment model are simulated on computers; the model that controls the real robots runs on the real robot. Couplings between the two robots are maintained the same as in conventional simulation. So the real and virtual robots interact with each 1 This may not be true if robot models have considered enough details of real robots. other in the same way as they do in the first step, although here the commutation actually happens across a network. The final step is the real system experiment, where all real robots run in a real physical environment. These robots use sensor/actuator interfaces (implemented as RTActivities) to drive real sensors and actuators. They communicate with each other in the same way as they do in the previous steps because the couplings between them are not changed through the process. Since all measurement results of this step come directly from the real system, they have the most fidelity. However, they are also most costly and time consuming to be collected. As described above, three types of DEVS Activities have been developed to act as sensors/actuators interfaces between a robot s decision-making model and the environment model. They are abstract Activity abstractactivity, real-time Activity RTActivity, and hardware-in-the-loop Activity HILActivity. These activities play different roles in different situations. An abstractactivity serves as a virtual sensor or actuator that is used by the decision-making model to interact with the environment model in simulation. An RTActivity is used in real execution to drive a real sensor or actuator. A HILActivity is employed in RIL simulation to drive a real sensor or actuator and also synchronizes with the environment model. It is important to note that in order to maintain the decision-making model unchanged, the corresponding abstractactivity, RTActivity, and HILActivity should maintain a same set of interface functions that are used by the decision-making model. This incremental simulation-based measurement process establishes an operational framework to measure and evaluate cooperative robotic systems. As the process proceeds, the flexibility (easy to setup different experiments) and productivity (time saving and cost saving) of the measurement decreases and the fidelity (loyal to the reality) of the measurement increases. 3 A ic Convoy System A robotic convoy system has been developed as a case study example to illustrate how the incremental measurement process works. This robot convoy system consists of an indefinite number of robots, saying N robots (N>1). These robots are in a line formation where each robot (except the leader and the ender) has a front neighbor and a back neighbor. The robots used in this system are car type mobile robots with wireless communication capability. They can move forward/backward and rotate around the center, and have whisker sensors and infrared sensors. One of the main goals of this convoy system is to maintain the coherence of the line formation and to synchronize robots movements. Synchronization means a robot cannot move forward if its front robot doesn t move, and it has to wait if its back robot doesn t catch up. To serve this purpose, synchronization messages are passed between a robot and its neighbors. To achieve
coherence of the line formation, the moving parameters of a front robot are passed backward to its immediate back robot. This allows the back robot to plan its movement accordingly based on its front robot s movement. Figure shows the model of this system. Each block represents a robot model. These robot models have input and output ports, which are used to receive/send synchronization messages as well as moving parameters. The model couplings are shown in Figure. Figure 3 shows a snapshot of a simulation of a convoy system with 3 robots within a field surrounded by walls (no obstacles inside). The simulations show that robots will not follow the same path of the leader robot. But they are able to go after their immediate front robots, thus forming a coherent team from the entire system point of view. FReadyIn n FReadyOut BReadyIn FReadyOut 3 FReadyIn BReadyOut BReadyIn FReadyOut FReadyIn BReadyOut BReadyIn 1 BReadyOut Figure : System model of the robotic convoy system During the convoy, the leader robot (1 in Figure ) decides the path of convoy. In this example, it moves straight forward if there is no obstacle ahead. Otherwise it turns right. All other robots conduct movement based on their IR sensory inputs and also the moving parameters passed from their front robots. Specifically, a robot first predicts where its front robot is and turns to that direction. It then moves forward (or backward) to catch its front robot. After that it may go through an adjust process to make sure that it follows its front robot. This adjust process is necessary because noise and variance exist during a movement so a robot may not reach the desired position/direction after a movement. During adjustment, a robot scans around until it finds its front robot. Then it sends out a synchronization message to inform its neighbors. Thus each robot executes a basic predict and turn move adjust inform routine in every cycle. The motivation to design the above control logic, especially the adjust process, is because there are significant uncertainty and inaccuracy in the movement of the mobile robots that we use. To model robots motion uncertainty, two noise factors: distance noise factor (DNF) and angle noise factor (ANF) are developed and implemented using random numbers. The DNF is the ratio of the maximum distance variance as compared to the robot s desired moving distance. The ANF is the ratio of the maximum angle variance as compared to the robot s desired moving distance. The sensor models model the sensing range ( cm) of IR sensors. Also, all IR sensory data are rounded to the closest multiplicity of (i.e., 1,, ), since this is how the real IR sensors work. The environment model is responsible to keep track of robots movement and provides sensory dada when robots need it. It includes TimeManager models and a SpaceManager model. A TimeManager models the time for a robot to complete a movement. The SpaceManager models the moving space, including the dimension, shape and the obstacles inside the field. It also keeps track of robots positions and moving directions. Such tracking is needed to supply robots with the correct sensory data. A robot s position is updated when the environment model gets moving command messages from the robot. Figure 3: Snapshot of robots in simulation /Experiment & Results Figure : Snapshot of RIL simualtion Following the incremental study process, simulations and experiments were carried out to measure the robotic convoy system described in Section 3. In particular, a movie [] was recorded for RIL simulation where two real robots are used. This movie demonstrates the coordination between real robots and robot models. Figure gives a snapshot from this movie, which shows how two real robots work together with robot models. In this movie, four robots (denoted by R, R1, R, and R3) are used, among which the second and third ones (R1 and R) are real robots. R1 uses virtual IR sensor to get sensory input from the environment model. R uses real IR sensor to sense its front robot (R1) and the real environment. As shown by Figure, this movie has two windows. The right window shows how two real robots move in the real world. The left window is the simulation window. It displays the movements of the entire convoy system, among which the second and third robots are the counterparts of the two real robots (R1 and R). This means the second and third robots movements in the simulation window are synchronized with the two real robots movements in the real world. Thus when a real robot moves backward because it is too close to its front robot, its counterpart in the simulation window moves backward too. Notice that in RIL simulation, a counterpart s position (direction) is updated based on the real robot s moving parameters. Due to noise, the actual distances that a real robot moves may be
different from the ones specified by the moving parameters. However, these errors are tolerable since wheel encoders are used by real robots to determine if a desired distance is reached. Quantitative results were also collected using the incremental measurement process that includes conventional simulation, RIL simulation, and real system experiment. Below we present these results from conventional simulation and RIL simulation using a convoy system with six robots. We also show the results from a real system experiment using two real robots..1 and RIL Setup Figure shows the convoy system that was studied using conventional simulation. The system includes six robot models, with being the leader robot. All robot models were equipped with virtual IR sensors and virtual motors to interact with an environment model. The environment model defines a cm 1cm open space surrounded by walls. The simulation stops when completes one circle. 3 1 Figure : Conventional simualtion with six robots Figure shows the setup of RIL simulation for the same system shown in Figure. The only difference is that the third and fourth robots, and 3 respectively, are real robots. uses virtual IR sensor to get sensory input, i.e., the distance to 1, from the environment model. 3 uses real IR sensor to get sensory input, i.e., the distance to, from the real environment. Both and 3 use motor HILActivities to move on a physical floor (without any obstacles). All other robots are models that use virtual IR sensors and virtual motors. In this system, because each robot (except ) s decision making is affected by its immediate front robot, this setup of RIL simulation divides the six robots into three categories. 1 Figure : RIL simualtion with six robots Category 1 (, 1): These two robot models exist in the virtual world. Their decision makings are not affected by the fact that two real robots are included in the simulation (1 needs to wait for the ready message from the real robot. However, that is only for synchronization purpose.). Because of this, it is expected the results collected in RIL simulation for these two models will be the same as those collected in conventional simulation. Category (, 3): These two real robots represent two different situations. 1) moves in the real world. However, its immediate front robot is 1, which is a model and is not affected by the two real robots. This means the sensory data, and hence the movement patterns, of in RIL simulation should be the same as (or very similar to) those in conventional simulation. However, due to noise and variance, the actual moving distances of in RIL simulation will be different from those in conventional simulation. ) 3 makes its decision solely based on the information from the real world, i.e., it receives information from and uses its IR sensors to check if it follows. Because 3 uses its real sensors and follows a real robot, it is expected the sensory data and the movement patterns of 3 in RIL simulation will be different from those in conventional simulation. Category 3 (, ): These two robot models exist in the virtual world. However, because s immediate front robot is 3 whose behavior changes when come to RIL simulation, it is expected the results collected in RIL simulation for will be different from the ones collected in conventional simulation. Similarly, s results change when come to RIL simulation. In fact, if there are more robots after, the results of those robots change too.. Measurement Metrics and Results Several measurement metrics were defined and simulation results were collected in one trial of conventional simulation and two trials of RIL simulation. To analyze these data, for each robot category described above, we pick up one robot and compare its conventional simulation results (referred to as simulation data) with its RIL simulation results (referred to as RIL data). These results show that, 1, s simulation data and RIL data are the same (similar); 3 s simulation data and RIL data are different; and s simulation data and RIL data are different too. Note that it is important to differentiate and 3 in the two simulations. In conventional simulation, they are models. In RIL simulation, they are real robots. Number of adjustment of each step In order to follow its front robot, a robot may go through an adjust process after each movement step. The number of adjustment of a robot thus indicates how smoothly this robot convoys. Figure 7 shows the number of adjustment for 1, 3, and. In the figure, the horizontal axis represents the movement steps; the vertical axis represents the number of adjustment. For example, the figure shows that 1 had adjustments at step in conventional simulation as well as in two RIL simulations. As we expected, these results show 1 had the same simulation data and RIL data. But 3 and s simulation data and RIL data are different.
for 1 Front IR Distances (cm) for 3 1 1 3 7 9 11 13 1 17 19 1 Front IR Distances (cm) 3 3 1 1 1 3 7 9 11 13 1 17 19 for 3 Front IR Distances (cm) for 3 1 1 1 8 1 3 7 9 11 13 1 17 19 Front IR Distances (cm) 3 3 1 1 1 3 7 9 11 13 1 17 19 3 1 for 1 3 7 9 11 13 1 17 19 Figure 7: Number of adjustment The results presented in Figure 7 show that for 3, the number of adjustment in RIL simulation at some steps is much larger than that in conventional simulation. For instance, the number of adjustment at step 17 in RIL trial is 11. However this value at the same step in conventional simulation is. This information, by comparing simulation data with RIL data, provides useful feedback to the designers, i.e., it indicates that the robot models may not model the real robots movement very well. On the other hand, from RIL simulation, we can see that even though 3 s number of adjustment becomes large at some steps, its IR distance data (presented in Figure 8) are still stable. For example, Figure 8 shows 3 s IR distance at step 17 in RIL trial is 3. This information, collected in RIL simulation using real robots, increases designers confidence about how the final real system is going to work. Notice that in both of these two cases, RIL simulation allows the designers to use only several, instead of all, real robots to gain the above knowledge. IR distance data of each step (after adjustment) We define two metrics to study how coherent this robotic system convoys. One of them is the front IR distance, which is the value returned from a robot s front IR sensor. Figure 8 shows the front IR distance of, 3, and. As can be seen, s IR distance data in simulation and in RIL are the same. But 3 and s IR distance data are different. More importantly, Figure 8 shows that even though 3 s IR distance changes, the change is within a boundary and does not accumulate as time proceeds (the desired IR distance is set to ). This information, together with the number of adjustment presented above, indicates the control model of this robot convoy system is robust. Front IR Distance (cm) 3 3 1 1 Front IR Distance (cm) for 1 3 7 9 11 13 1 17 19 Figure 8: Front IR distance (cm) Coherence data The coherence data is another metric that is defined to study how coherently robots convoy. It calculates the difference between a robot s actual position and its desired position (based on its front robot). The formula used for the calculation can be found in [1]. Figure 9 shows the coherence data for, 3, and. Similar to the reason explained above, coherence data does not change from simulation to RIL, while 3 and s coherence data change. Coherence Data Coherence Data Coherence Data 3 1 7 3 1 7 3 1 Coherence Data for 1 11 1 31 1 1 1 71 81 91 11 111 11 Steps Coherence Data for 3 1 11 1 31 1 1 1 71 81 91 11 111 11 Steps Coherence Data for 1 11 1 31 1 1 1 71 81 91 11 111 11 Steps Figure 9 : Coherence data of robots Figure 1 illustrates the average coherence of the five robots (except, whose coherence is ). It clearly
shows that although the average coherence data are different in simulation and in RIL, they are still consistent. Similar consistency between simulation data and RIL data can also be seen in Figure 9, for example, the coherence data of 3. This type of consistency among simulation data, RIL data, and real experiment data (presented next) conveys two important messages to us. First, it provides some level of validation to the simulation models used in conventional simulation. Secondly, it justifies the model continuity methodology and the incremental simulation-based measurement process that we applied to develop this robotic convoy system. Average Coherence Data 3 1 Average Coherence Data 1 11 1 31 1 1 1 71 81 91 11 111 11 Steps Figure 1 : Average coherence data.3 System Experiment and Results system experiments were also carried out by applying the same decision-making models to two real robots. In these experiments, two real robots were placed in a cm 1cm open field surrounded by walls (boxes). Both robots use real IR sensors and motors to sense and move within the real environment. A movie for one of these experiments can be found at []. Measurement data were also collected. Figure 11 shows the number of adjustment and front IR distance for the second real robot in three trails of experiments. As expected, these results demonstrate similarities to the results from conventional simulation and RIL simulation. For example, by calculating the average IR distance data for 3 in conventional simulation, RIL simulation, and real execution, we have 3.cm,.3cm, and 1.3cm respectively. Front IR Distance (cm) 1 1 8 3 3 1 1 in Execution 1 3 7 8 9 1 11 1 13 1 1 1 Front IR Distance (cm) in Execution 1 3 7 8 9 1 11 1 13 1 1 1 Figure 11: system experiment results Trial1 Trial Trial3 Trial1 Trial Trial3 Discussion and Future Work Both the measurement data, such as the front IR distances, and the recorded movies show the continuity of evolving this robotic convoy system from conventional simulation, to RIL simulation, and then to real system execution. This is because a model continuity methodology is applied where the same control models are maintained through different stages of system development. The quantitative results from simulations/experiments demonstrate the feasibility of carrying out an incremental simulation-based design process by gradually bringing real system components into the design until the system evolves into its final form. This capability is especially useful for large-scale complex systems. It provides an operational framework to support simulation models and real system components to work together for system-wide test and measurement. RIL simulation, as a major step in this process, effectively bridges the gap between conventional simulation and real system experiment. As a prelude for actual implementation, RIL simulation brings simulationbased study closer to the reality and increases the confidence that the final system will work as designed. To summarize, this research affords several flexibilities to testing and measurement of distributed robotic systems: a) flexibility to study real robots in a virtual environment; b) flexibility to study models based on inputs/interactions from real robots; c) flexibility to study a large-scale multirobot system using combined real and virtual robots. We note that the incremental design process presented in this paper should not be limited to robotic convoy applications only. As one future research task, we plan to apply this process to other application areas. In addition, we plan to develop experimental frames for each step and add more complexity to the robotic system to check how effectively this process can handle more complex situations. References [1] X. Hu, B. P. Zeigler, Measuring Cooperative ic Systems Using -Based Virtual Environment, Performance Metrics for Intelligent Systems Workshop, August [] X. Hu, A -based Software Development Methodology for Distributed -time Systems, Ph.D. Dissertation, University of Arizona, [3] B. P. Zeigler, H. Preahofer, T. G. Kim, Theory of Modeling and, New York, NY, Academic Press,. [] Movie http://www.cs.gsu.edu/~cscxlh/ril.swf, SWF movie file, playable using Internet Explore [] Movie http://www.cs.gsu.edu/~cscxlh/twos.swf, SWF movie file, playable using Internet Explore