On the Development of Cooperative Behavior-BasedMobile ManipulatorsBruno S. Pimentel Guilherme A. S. Pereira M´ario F. M. CamposVERLab – Laborat´orio de Vis˜ao Computacional e Rob´otica,Departamento de Ciˆencia da Computac¸˜ao,Universidade Federal de Minas Gerais, Belo Horizonte, Brazil[brunosp, gpereira, mario]@dcc.ufmg.brABSTRACTWe present an approach to the problem of cooperative objectcarrying, to be performed by a group of mobile robots con-trolled with a behavior-based architecture. Inherent char-acteristics of this method such as simplicity, rapid devel-opment, robustness, low memory and processing require-ments have allowed us to implement a behavioral controlsystem in rather simple, low-processing-power mobile plat-forms. Robot coordination may be achieved through eitherimplicit or explicit communication. Although implicit com-munication can make the system more robust to faulty com-munication environments, we show that the use of the ex-plicit form can avoid some undesirable system locks. Ex-perimental results where two robots carry a relatively largebox in an environment cluttered with both easily and hardlyavoidable obstacles show the flexibility and effectiveness ofthe proposed architecture.1. INTRODUCTIONMobile-robot control systems have been implemented in var-ious ways, ranging from purely deliberative to purely reac-tive. Deliberative reasoning has its bases on traditional ar-tificial intelligence, and relies on relatively complete worldmodels to predict the outcome of robot actions, so as tooptimize its performance for some given criteria. This, how-ever, makes those systems heavily dependent on symbolicrepresentational world models, which must be continuouslyupdated with incoming sensor data in order to allow robotoperation in dynamic domains. Although allowing more re-liable and robust construction of suitable solutions, deliber-ative control is harder to implement efficiently. On the otherhand, reactive robotic systems tightly couple perception toaction – typically in the context of motor behaviors – to pro-duce real-time robotic response in dynamic, unstructuredenvironments, without the use of abstract representationsor time history [2]. Reactive systems were the basis for thedevelopment of recent behavior-based architectures, whichrely heavily on sensing without constructing potentially er-roneous global world models, thus being better suited tosituations where the real world cannot be accurately char-acterized. Although optimal (or even viable) actions maynot be produced – specially in more complex tasks – lowercomputational requirements usually result in real-time robotresponse for incoming sensor data.In the context of cooperative robotics, several complex sit-uations involving coordination, communication and interac-tion between multiple agents may arise. The main goal of acooperative system is to carry out a task improving its over-all performance or even allowing it to be completed at all.Loosely coupled cooperation arises in those tasks where themain objective does not require close interaction betweenrobots and the goal can be distributed amongst the team,using a strategy similar to divide-and-conquer. For instance,if the task at hand is to explore an unknown environment,it could be more efficiently accomplished by partitioning thearea to be covered among the team members. The failureof one or more robots should not compromise the main goal(but probably increase the total exploration time), providedthat there were enough units to finish the job. On the otherhand, if a task can only be completed through the interac-tion of a minimum number of robots, it is called a tightlycoupled task. For example, a single robot is not able tocarry a heavy, large box if it doesn’t have enough power orgrasping ability [7]. Also, a group of mobile robots acting ascooperative predators that must capture a prey is anotherexample of a tightly coupled task since a single robot aloneis unable to confine the prey [8]. Given the team capabili-ties, these tightly coupled tasks could only be satisfactorilycompleted by close interaction of a number of robots.Reactive systems have been extensively implemented in manysimple loosely coupled tasks. On the other hand, deliber-ative control seems to be better suited to tightly coupledtasks, due to their inherent complexity in coordinating robotinteraction. However, the simplifying characteristics of thefirst motivate us to apply the reactive control paradigm onthe development of a behavior-based architecture that al-lows a group of cooperative mobile manipulators to carry alarge object, while navigating in an unknown, unstructuredenvironment.Cooperative transport is a task found in many real-worldsituations. Ants of various species, for example, perform itin order to move back to their nest large preys otherwise im-possible for a single ant to retrieve. As pointed by Kube andBonabeau in [11], the coordination of ants collective trans-port seems to occur through the item being transported,since a movement caused by one ant modifies the stimuliperceived by its teammates, which consequently produceschanges in their position and orientation. Like the ants,our robots sense the movement of each other through theobject being carried. Basic behavioral building blocks weredesigned and combined to produce actions of targeted nav-igation, obstacle avoidance, and object grasping control.The remaining of the paper is structured as follows: Sec-tion 2 discusses some of the applications of the behavior-based paradigm and previous approaches to the coopera-tive object-carrying task. Section 3 describes the problemtackled in this work. Methodology and system architectureare both presented in Section 4, while Section 5 validatesthe proposed approach through real-world experiments. Fi-nally, the main contributions are summarized and discussedin Section 6.2. RELATED WORKBehavior-based cooperative systems have been successfullyimplemented to perform a wide range of loosely coupledtasks [16, 3]. Formalized behavior-based software archi-tectures have also been developed to allow heterogeneousmulti-robot cooperation [13]. Yet, Sugar et al. [18] observedthat it is not clear whether the behavior-based methodol-ogy could be applied to controlling grasp forces in order tohold and transport objects [7], which is an example of atightly coupled task. Although Matari´c et al. [12] and Kubeand Bonabeau [11] have implemented cooperative