An Experimental Testbed for Robotic Network Applications

Transcription

1 An Experimental Testbed for Robotic Network Applications Donato Di Paola, Annalisa Milella, and Grazia Cicirelli Abstract In the last few years, multi-robot systems have augmented their complexity, due to the increased potential of novel sensors and actuators, and in order to satisfy the requirements of the applications they are involved into. For the development and testing of networked robotic systems, experimental testbeds are fundamental in order to verify the effectiveness of robot control methods in real contexts. In this paper, we present our Networked Robot Arena (NRA), which is a software/hardware framework for experimental testing of control and cooperation algorithms in the field of multi-robot systems. The main objective is to provide a user-friendly and flexible testbed that allows researchers and students to easily test their projects in a real-world multi-robot environment. We describe the software and hardware architecture of the NRA system and present an example of multi-mission control of a network of robots to demonstrate the effectiveness of the proposed framework. 1 Introduction In these last years, multi-robot systems are being widely investigated, since they offer better performances than single robot systems in challenging applications, such as exploration of hostile environments, terrain mapping, space and rescue operations [1],[2],[3],[4]. Due to the intrinsic difficulty in developing and assessing the performances of multi-agent control algorithms in real contexts, much work has been carried out only analytically or in simulation [5],[6],[7]. Advanced simulation tools for multi-robot systems have been also developed, and are available as open source tools (e.g., Stage and Gazebo) or as commercial products (e.g., Webots, Microsoft Robotic Studio etc.). Donato Di Paola, Annalisa Milella, Grazia Cicirelli Institute of Intelligent Systems for Automation (ISSIA), National Research Council (CNR), via G.Amendola 122/D, Bari, Italy, {dipaola, milella, grace}@ba.issia.cnr.it 1

2 2 Donato Di Paola, Annalisa Milella, and Grazia Cicirelli Nevertheless, for complete analysis and testing, and to provide additional insights to theory, real world experiments are generally desirable. On the other hand, experimentation in robotics for comparison of different methods is often difficult due to the variety of robotic platforms and environments that may be encountered. Therefore, the development of experimental testbeds is a key issue for the design of effective multi-robot control systems. These testbeds can support all stages of the design process, including algorithm development and testing, conventional simulation, robot-in-the-loop simulation, and real robot control, in an integrated way. An example of platform for conducting, evaluating and analyzing experiments in robotics, named the Teleworkbench, can be found in [8]. In this paper, we describe our experimental set-up, referred to as Networked Robot Arena (NRA) that allows us to monitor, control, and test realistic behaviours of relatively simple multi-robot systems. This set-up includes both hardware and software components. The main contribution of the work is the development of a low cost, robust, flexible, and scalable software/hardware system, which uses desktop mobile robots with a vision-based identification and localization system, and a multi-mission control framework. The control architecture is a hybrid distributed architecture in which the high-level multi-mission control is performed by a software module with a number of controllers, each one in charge of managing a sub-network, while the low-level operations are implemented on board each robot. The proposed framework can serve as a middle ground between a totally software-based simulation tool and a full-scale real-world implementation. From the one hand, it allows one to encompass real-world problems difficult to model and manage in a virtual environment, such as errors in robot identification and localization, or interactions with other objects. On the other hand, it is based on user-friendly and flexible tools, which can be used by both researchers and students to easily test their projects. 2 Overview of the Networked Robot Arena In this section, we provide an overview of the NRA. The main goal of the proposed system is to set up a hardware and software framework to assist the development and testing process of robotic network applications for educational and research purposes. The NRA has the following characteristics, which are typically desired for education and research testbeds: open software and hardware platform; desktop mobile robots; low cost; robust, flexible and scalable software; user friendly software development and maintenance. In order to satisfy these requirements, the NRA features the logical structure shown in Figure 1. The main components of the system are the arena, i.e., the physical environment where the robots operate, and the workstation, i.e., the computer in charge of communicating with the robots and controlling most of software modules.

3 An Experimental Testbed for Robotic Network Applications 3 The arena is a bounded and controlled environment in which the robots, wireless connected, can move to perform the desired tasks. In order to satisfy the small size requirements, in our implementation, the arena is a desk with boundaries and movable objects (i.e., small wooden items), which can be arranged to create walls and divide the environment in different zones. A camera is placed high above the arena to monitor positions and actions of the robots over time. This camera sends images to the workstation for elaboration, through a USB port. The output of the image processing algorithm consists of a set of robot IDs and their positions in the arena, which are fed into the control system of the network. A short-range wireless connection is adopted to link the robot network with the workstation. The latter runs a software architecture, which is in charge of: managing the connections between the workstation and the network of robots, using appropriate communication libraries; processing the images acquired by the overhead camera; controlling the network with high-level decision making algorithms; communicating the control decisions to the robots. Indeed, control is not totally centralized in the control module of the architecture, but it is distributed between the workstation and the network of robots. Furthermore, no a priori knowledge about the environment configuration is required by the control system. In order to facilitate the use of the testbed by researchers and students, MATLAB is employed as software development environment. In the next sections, each part of the system architecture is described in more detail. 2.1 Robots and Hardware Infrastructure As discussed above, two of the design requirements of the NRA are the desktop-size dimensions of the robots and the possibility to have an open hardware platform. Keeping in mind these requirements, we analyzed different solutions. First, we selected three robots as possible candidates for the NRA platform: the K-Team Khepera robot [9], the Lego Mindstorms NXT, and the E-Puck robot. Khepera is Fig. 1: Schematic representation of the Networked Robot Arena (NRA).

4 4 Donato Di Paola, Annalisa Milella, and Grazia Cicirelli a small size robot, which can be equipped with various sensors, like infra-red proximity and ambient light sensors, cameras, and ultrasonic sensors, but it is relatively expensive. The Lego Mindstorms NXT is a modular platform with a rich set of sensors; however, it is not an open platform, and it is not suitable for desktop size robotic networks. Finally, we chose the E-Puck robot (see Figure 2(a)), because of its small size, low cost and open hardware characteristics [10]. The E-puck is a desktop size (7.0 cm diameter) differential wheeled mobile robot. It was originally designed as an education tool for engineering and robotics. It is equipped with a powerful Microchip dspic (i.e., dspic 30F6014A) microcontroller, and the following sensors and actuators: eight infrared (IR) proximity sensors, placed around the body; a 3D accelerometer that can be used to measure the inclination and acceleration of the robot; three microphones, to localize source of sound by triangulation; a CMOS colour camera ( ), mounted in the front of the body; two stepper motors for differential drive; a speaker, connected to the audio module; a set of LEDs placed on the body to provide a visual interface to the user; a set of communication interfaces (i.e., a Bluetooth radio link, a RS232 serial interface, and an infrared remote control receiver). The hardware infrastructure of the NRA testbed consists of a bounded arena (60 cm 110 cm) with a wireless communication network, a USB camera, and a high performance workstation (see Figure 2(b)). The camera consists of a USB colour device, placed at about 150 cm above the arena. It provides robot identity (a) (b) Fig. 2: Robots and Hardware Infrastructure: (a) close up of an E-Puck robot inside our arena; (b) the Networked Robot Arena.

5 An Experimental Testbed for Robotic Network Applications 5 and position information for the control system, and can be also used as groundtruth basis for experimental validation of the algorithms. For the given arena and robot size, an image resolution of is sufficient to detect and track all the robots. The workstation consists of a high performance computer, equipped with an Intel Core 2 vpro microprocessor and 4 Gb RAM, and is used for control, visual data processing, and connection with the network of robots. Since we need short-range communication among the agents, and between the agents and the workstation, we use a Bluetooth wireless network. Due to the small amount of data passed through the network and the hardware and software module embedded within the E-Puck, the Bluetooth communication protocol revealed a good choice. 2.2 Software The main goal of the NRA software architecture is to provide a distributed mission control system, able to manage all the activities of the robotic network. We assume that the network has to execute a number of missions. Each mission consists of a predefined sequence of tasks, connected by logical rules. Activation and completion Fig. 3: Layout of the software architecture.

6 6 Donato Di Paola, Annalisa Milella, and Grazia Cicirelli of missions and tasks are triggered by input and output events, respectively. In order to execute each task, we assume that a set of behaviours, i.e. simple reactive actions, is properly activated on board each robot. The layout of the software architecture is shown in Figure 3. It consists of the following main parts: the high-level control, communication, and localization and identification modules, running on the workstation, and the execution module running on each robot. All components are executed on a discrete-time basis: state updates, event detection, and commands are all processed at the beginning of each sample time. The Controller implements high-level control functionalities, monitoring mission execution and performing task assignment. In this module, one or more controller can be run. Each controller, called leader, is in charge of controlling a part of the network, in terms of missions and tasks. More specifically, each leader controls the execution of the missions in progress (e.g., guaranteeing the satisfaction of precedence constraints), sends the configuration information about the tasks that must be started on each robot, and receives information about the completed tasks. Each leader performs its work using a discrete-event model of the domain, a task execution controller, and a conflict resolution strategy, which will be described in the following section. Obviously, if only one leader exists, the control of the network can be viewed as a centralized system. The Executor performs the control at the lower level. This component is implemented on each robot through a set of tasks and a set of behaviours needed to accomplish a given goal. For each task, multiple behaviours can be executed at the same time. Each behaviour computes an output and when multiple behaviours are active at the same time, a predefined behaviour arbitration algorithm (e.g., subsumption [11] or cooperative methods [12]) can be used to obtain the final control signal. Currently, in our system, a subsumption mechanism is implemented. At the end of a task, the completion flag and the results are sent to the upper level. Communication between the Executor and the Controller is guaranteed by the Communication module. This module provides all high-level control commands from the leaders to the other robots in the network. In particular, in this module, the proper communication libraries are loaded, according to the specified robot. Moreover, the communication module is connected to the vision based localization and identification module that provides, for each robot, position and identity information, used to accomplish an assigned task and for behaviour control. It is worth to note that, although the main objective of this work is that of providing an experimental platform, the proposed system may also be used as a simulation tool by substituting the Executor with an additional Simulation module. The latter, which uses the same interface with the Communication module as the Executor, simulates the behaviour of each robot in a virtual environment, providing robot positions and simulating task execution. In the rest of this section, further details concerning the hybrid control of the network, as well as the localization and identification module, are provided.

7 An Experimental Testbed for Robotic Network Applications Robot Network Control As mentioned above, the control of the network is distributed among the components of the architecture. The overall management of the robotic network is based on a hybrid control framework: the global network activities are modeled and controlled as a Discrete Event System (DES) and the low-level behaviours, for each robot, are considered as continuous-time control algorithms trigged by the DES supervisor. Moreover, the global model of the network is distributed among a set of leaders. Each leader can control all activities or a sub-set of them in the network. The activities domain of the network is modeled as follows: each robot must execute a set of missions; each mission is a set of atomic tasks, which may be subject to priority constraints (i.e., the end of one task may be a necessary prerequisite for the start of some other ones). To execute each task, the robot must perform a predetermined sequence of behaviours (e.g. GoTo, AvoidObstacle, WallFollowing, Wandering, etc.). Each behaviour runs until an event (either triggered by sensor data or by the supervisory controller) occurs. Two or more behaviours can be active simultaneously. In such a case, the behaviour arbitration algorithm, on board the robot, computes the appropriate command to be sent to robot actuators. Missions may be triggered at unknown times, while the number and type of tasks composing a mission is known a priori. Since different missions can run simultaneously, different tasks may require the same subset of robots. In these cases, conflicts between the robots have to be handled using an efficient task allocation policy in order to avoid deadlock situations. The overall control scheme, proposed in this work, is based on the Matrix-based Discrete Event modeling and control Framework (MDEF) [13], a tool for modelling and analyzing complex interconnected DES with shared resources, for routing decisions, and managing dynamic resources. More precisely, in this work a partial decentralization of the MDEF is implemented. The set of leaders are considered as a group of centralized MDEFs working on separate parts of the network [14]. All the MDEF builds over a set of key matrices and vectors. Although this presentation is mainly intended for self-containment purposes, it is obviously not possible to provide all the details of the formalism in the limited space available here. Reader therefore is referred to [14] for further details. Suppose that the RN is composed of V = {v 1,...,v n } robots, which have to accomplish M = {m 1,...,m l } missions. Each mission is, in turn, viewable as a sequence of primitive tasks (e.g., reach the target, take a picture of the target, measure the temperature, collect the target, etc.), which requires some low-level behaviors and may be subject to precedence constraints (e.g., the target can be collected after it has been found). Thus, T = {t 1,...,t m } denotes the set of all possible tasks for the RN. The execution of tasks is governed by a set of rules. More specifically, each rule specifies all the preconditions for tasks execution, and the resulting consequences (postconditions). All rules for all missions in M are defined by the set X = {x 1,...,x q }. Thus a mission m j is associated to a specific set of rules X m j X. The set of all possible input events (a sensory input, a user command, etc.) is indicated as U = {u 1,...,u p }. One or more inputs can be considered as

8 8 Donato Di Paola, Annalisa Milella, and Grazia Cicirelli preconditions for a generic mission m j, thus these event are element of the set U m j U. Besides tasks and rules, generic outputs can be considered as postconditions. All possible outputs in the network are indicated by the set Y = {y 1,...,y o }. Briefly, the MDEF controls the execution of tasks by checking the status of the rules at each time sample. Each rule whose preconditions are all true in the current time sample is fired, i.e., all the actions described in the consequent part of the rule are triggered. The result of the logical preconditions evaluation provide the information needed to properly trigger tasks. Since conflicts in the assignment of tasks can occur, before starting the execution of task a multi-agent task allocation algorithm is performed, such as the Look-Ahead strategy proposed in [15]. After the task allocation step, postconditions are fired properly and robots start assigned tasks and eventually output events are produced Robot Identification and Localization Robot identification and localization is performed using a vision-based tracking system. Indeed, E-Puck robots are equipped with encoders, and therefore odometry could be adopted for robot location estimate. Nevertheless, we use an external camera, since it provides an absolute positioning system, thus avoiding error accumulation problems. In addition, based on visual information, the robot identification issue can be easily solved. We developed an algorithm that is able to simultaneously estimate the global positions of all the robots in the arena, determining, at the same time, which location belongs to which robot. We call this module Simultaneous Localization and Identification (SLI). The output of the SLI module is fed into the control system. It is assumed that each robot is equipped with a circular coloured hat (see Figure 2(b)). Different colours are used for the different robots. Each hat serves as (a) (b) Fig. 4: Results of the SLI module: (a) wandering behaviour; (b) wall-following behaviour.

9 An Experimental Testbed for Robotic Network Applications 9 an artificial landmark for robot detection and identification. Specifically, the SLI includes a semi-automatic learning procedure that allows us to learn and store a model for each landmark; then, based on this model, it performs robot localization and identification through a model matching technique. The learning phase consists of the following steps. First, various pictures of each robot in the arena are acquired by the overhead camera; then, in each picture a polygonal area containing the robot is selected manually, and shape and colour information are extracted by analyzing the pixels of the selected area. Shape information is represented by using the radius of the minimum enclosing circle for the selected polygonal area. Colour is defined, instead, by a three-dimensional vector containing the median of each color component in the HSV color space. The procedure is repeated for all the pictures, and median values for both the radius and the color vector are computed and stored. Once the model of each robot has been learned, it can be used for detecting and identifying all the robots in the arena. The detection phase is based on a Circular Hough Transform, which allows one to detect all the circular objects within a specified radius interval based on the results of the model learning phase. Then, the pixels internal to each detected circle are analyzed for robot identification, using colour information. Specifically, the median of hue, saturation and value of the pixels internal to each circle are computed and are stored in a 3D color vector, which is then compared with the available robot models. A circle is finally recognized as being one of the robots if the Euclidean distance between its associated color vector and the closest robot color model is less than an experimentally determined threshold. At this step, the position of each robot is expressed as pixel coordinates of the centre of the circle in the image reference frame. A camera calibration procedure performed once, after camera installation, allows us to transform the pixel position in a real-world position. In order to improve the performance of the algorithm in terms of speed and accuracy of both detection and identification phase, we implemented a Kalman filter-based robot tracking. As an example, Figure 4 shows the result of the SLI module for four robots in two different experiments. Specifically, in Figure 4(a) the robots are tracked while wandering in the arena; in Figure 4(b), each robot carries out a wall-following task. 3 Experimental Results In this section, we show the results of an experimental test performed using the setup described in the previous sections. In this experiment, the network evolves from a single-leader configuration to a multiple-leader configuration, in order for the robots to perform exploration tasks in different parts of the arena starting from a rendez vous position. In the various phases of the test, the leader robots can be distinguished from the other ones, as they are green-lighted. In Figure 5(a), the four robots form a unique network with one leader (i.e., the green-lighted robot with the red hat). Figure 5(b) shows the trajectories of the four robots while going to

10 10 Donato Di Paola, Annalisa Milella, and Grazia Cicirelli the rendez vous position. Successively, the network splits into two sub-networks of two robots each (Figure 5(c)). Both networks have their own leader (i.e., the greenlighted robot with the blue hat and the green-lighted robot with the red hat). The two teams reach opposite sides of the arena to explore different areas. Figure 5(d) shows the tracked trajectories of the robots after the splitting of the network. The graph in Figure 6 shows the time trace of the experiment. This graph can be easily obtained by the saved data of the logging system. In particular, in this figure, we show: the network configuration, i.e. the single leader network (Network AB), and the two sub-networks (Network A and Network B); the missions in progress; the robots assigned to each mission and their network; the activation of tasks. For all these information a line representing high (activation) or low (deactivation) state is depicted. In Table 1, the correspondences among missions, robot networks, robots and tasks are reported for better understanding the graph in Figure 6. It can be noticed that Mission 1, which involves all robots (i.e., Network AB), starts at sampling instant 7. This mission is accomplished at time 33, when each robot in the network terminates Task 1, that is, the GoToGoal task, in order to reach the rendez vous (a) (b) (c) (d) Fig. 5: (a)-(b) Rendez vous of the networked robots with a unique leader; (c)-(d) splitting of the network into two sub-networks with two different leaders to explore opposite sides of the arena

11 An Experimental Testbed for Robotic Network Applications 11 Table 1: Correspondences among missions, robot networks, robots, and tasks Mission ID Network Name Robot ID Task ID 1 AB 1,2,3, B 2,4 1,3 3 A 1,3 1,2 point at the centre of the arena. After Mission 1, the network splits into the two sub-networks A and B. As we note, network B starts Mission 2 before network A at the time instant 33. Obviously, the robots #2 and #4, which are members of network B, become busy when Mission 2 starts, performing, first, Task 1 and, then, at time 47, Task 3 (WallFollowing). At the instant 36, Mission 3 starts involving robots #1 and #3. Network A, first, performs Task 1 and, then, Task 2 (ExploreEnvironment). At completion of Mission 3, at time 68, the robots of network A become available and all the tasks they are involved in terminate. Finally, at time 74, Mission 2 is accomplished, robots #2 and #4 are released, and all tasks in the whole network are successfully terminated. Fig. 6: Time trace graph for the experiment of Figure 5.

12 12 Donato Di Paola, Annalisa Milella, and Grazia Cicirelli 4 Conclusions We described an experimental set-up for multi-robot applications. The small sized, relatively simple robots E-Puck are utilized. A USB camera connected to a high speed PC is adopted for robot monitoring and global positioning. Image processing provides the positions of the robots to the control algorithms. The latter are based on discrete event framework, for the high level control, and a task and behaviour-based distributed control framework at the lower level. We believe that this test bed is a useful experimental facility, which can be employed for testing multi-robot coordination and control algorithms in graduate and undergraduate courses as well as for research purposes. References 1. Burgard, W., Moors, M., Stachniss, C., and Schneider. F., Coordinated multi-robot exploration. In IEEE Transactions on Robotics, 21(3). 2. Fierro, R., Das, A.,, Spletzer, J., Esposito J., Kumar, V., Ostrowski, J.P., Pappas, G., Taylor C.J., Hur Y., Alur, R., Lee, I., Grudic G., Ben Southall, B., A Framework and Architecture for Multi-Robot Coordination, In The International Journal of Robotics Research, 21(10-11), pp Kumar, V., Rus, D., and Singh, S., Robot and sensor networks for first responders, In IEEE Pervasive Computing, vol. 3, no. 4, pp , Oct. 4. Vincent, R., Fox, D., Ko, J., Konolige, K., Limketkai, B., Morisset, B., Ortiz, C., Schulz, D., and Stewart, B., Distributed multirobot exploration, mapping, and task allocation. In Annals of Mathematics and Artificial Intelligence, Springer Netherlands. 5. Pugh, J. and Martinoli, A., Multi-robot learning with particle swarm optimization. In Proceedings of the Fifth International Joint Conference on Autonomous Agents and Multiagent Systems. 6. Lerman, K., Chris Jones, C., Galstyan, A., Matarc, M.J Analysis of Dynamic Task Allocation in Multi-Robot Systems, International Journal of Robotics Research, 25(3), pp Viguria, A., Maza, I., and Ollero, A., SET: An algorithm for distributed multirobot task allocation with dynamic negotiation based on task subsets. In Proc. of 2007 IEEE International Conference on Robotics and Automation, Roma, Italy. 8. Werner, F., Rckert, U., Tanoto, A, Welzel, J The Teleworkbench A Platform for Performing and Comparing Experiments in Robot Navigation, IEEE International Conference on Robotics and Automation (ICRA), Anchorage, AK, USA 9. Mondada, F., Franzi, E. and Guignard, A., The Development of Khepera. In Experiments with the Mini-Robot Khepera, pp Mondada, F., Bonani, M., Raemy, X., Pugh, J., Cianci, C., Klaptocz, A., Magnenat, S., Zufferey, J.-C., Floreano, D. and Martinoli, A., The E-Puck, a Robot Designed for Education in Engineering. In Proceedings of the 9th Conference on Autonomous Robot Systems and Competitions, 1(1) pp Brooks R.A., Intelligence without Representation. In Artificial Intelligence, Vol.47, pp Payton D.W., Keirsey D., Kimble D.M., Krozel J. and Rosenblatt J.K., Do whatever works: A robust approach to fault-tolerant autonomous control, In Applied Intelligence, 2 (3), pp

13 An Experimental Testbed for Robotic Network Applications Tacconi, D., Lewis, F., A new matrix model for discrete event systems: application to simulation. In IEEE Control System Magazine, Volume 17, Issue 5, pp Di Paola, D., Gasparri, A., Naso, D., Ulivi, G., and Lewis, F. L., Decentralized Task Sequencing and Multiple Mission Control for Heterogeneous Robotic Networks, In Proc. of IEEE International Conference on Robotics and Automation. 15. Di Paola, D., Naso, D., and Turchiano, B., A Heuristic Approach to Task Assignment and Control for Robotic Networks, In Proc. of the IEEE International Symposium on Industrial Electronics (ISIE).