【正文】 m with this approach is that the amount of munication needed between the master and the modules will limit its scalability. Another problem is the need for a central controller, since it gives the system a single point of failure. If there is no master it is suggested that the modules can be assumed to be synchronized in time and each module can execute its column of actions openloop. However, since all the modules are autonomous it is a questionable assumption to assume that all the modules are and can stay synchronized. In order to use the gait control table each module needs to know what column it has to execute. This means that the modules need IDs. Furthermore, if the configuration changes or the number of modules changes the table has to be rewritten. Shen et al. [17] propose to use artificial hormones to synchronize the modules to achieve consistent global lootion. In earlier versions of the system a hormone is propagated through the selfreconfigurable system to achieve synchronization. In later work the hormone is also propagated backward making all modules synchronized before a new action is initiated [16,18]. This synchronization takes time O(n), where n is the number of modules. This slows down the system considerably, because it has to be done before each action. Also, the entire system stops working if one hormone is lost. This is a significant problem, because a hormone can easily be lost due to unreliable munication, a module disconnecting itself before a response can be given, or a module failure. In fact, the system has n points of failure which is not desirable. The earlier version is better in this sense, but still performance remains low because a synchronization hormone is sent before each action. Butler et al. [4] propose a method inspired by cellular automata. In their approach modules respond to state changes of neighbor modules. Their approach is a bottomup approach related to ours, but in cellular automata there is no concept of time only of sequence. Timing is important in lootion, because it is the key to produce smooth and lifelike lootion and avoid jerky lootion. In our system all modules repeatedly go through a cyclic sequence of joint angles describing a motion. This sequence could e from a column in a gait control table, but in our implementation the joint angles are calculated using a cyclic function with period T . Every time a module has pleted a specified fraction d of the period a message is sent through the child connectors. If the signal is received the child module resets its action sequence making it delayed d pared to the parent. This way the actions of the individual module are decoupled from the synchronization mechanism resulting in a faster and more reliable system. Furthermore, there is no need to make changes to the algorithm if the number of modules changes. 3. Rolebased control We assume that the modules are connected to form a tree structure, that a parent connector is specified, and that this connector is the only one that can connect to child connectors of other modules. Furthermore, we assume that the modules can municate with the modules to which they are connected. The algorithm is instantiated by specifying three ponents. The first ponent is a cyclic action sequence A(t), where t ∈ [0 : T ]. T is the second ponent that needs to be specified and is the period of the action sequence. A(t) describes the actions that each module repeats cycle after cycle. In our implementation A(t) returns joint angles to control the two degrees of freedom of the CONRO module, but the action sequence could also be used to trigger different behaviors at different times during a cycle. The third ponent is a delay d. This delay specifies the fraction of a period the children’s action sequences are delayed pared to their parents. The skeleton algorithm looks like this: t=0 while(true) { if(t=d)then signal child modules if parent signals then t=0 perform action A(t) t=(t+1) modulus T } Ignoring the first two lines of the loop, the module repeatedly goes through a sequence of actions parameterized by the cyclic counter t. This part of the algorithm alone can make a single module repeatedly perform the specified sequence of actions. In order to coordinate the actions of the individual modules to produce the desired global behavior the modules need to be synchronized. Therefore, at step t = d a signal is sent through all child connectors. If a child receives a signal it knows that the parent is at t = d and therefore sets its own step counter to t = 0. This enforces that the child is delayed d pared to its parent. From the time the modules are connected it takes time proportional to d times the height of the configuration tree for all the modules to synchronize. To avoid problems with uncoordinated modules initially we make sure the modules do not start moving until they receive the first synchronization signal. After the startup phase the modules stay synchronized using only constant time. 4. Experimental setup To evaluate our algorithm we conducted several experiments using the CONRO (CONfigurable RObot) modules of which one is shown in Fig. 1. The CONRO modules have been developed at USC/ISI [5,9]. The modules are roughly shaped as rectangular boxes measuring 10 cm cm cm and weigh 100 g. The modules have a female connector at one end and three male connectors located at the other. Each connectorhas a infrared transmitter and receiver used for local munication and sensing. The modules have two controllable degrees of freedom: pitch (up and down) and yaw (side to side). Processing is taken care of by an onboard Basic Stamp 2 processor. The modules have onboard batteries, but these do not supply enough power for the experiments reported here and therefore the