【正文】 col [1] implemented on a dedicated IC (SMC COM20200). ARCNET is a deterministic tokenpassing work scheme which avoids 7 work collisions and guarantees each node its time to access the work. Blocks of information called packets may be sent from any node on the work to any one of the other nodes, or to all nodes simultaneously (broadcast). Each node may send one packet each time it gets the token. The maximum work throughput is 5Mb/s. The first node of the work resides on the host interface card, as is depicted in Figure 6. In addition to a VME address decoder, this card contains essentially the same hardware one can find on a module motherboard. The munication between the VME side of the card and the ARCNET side occurs through dualport RAM. There are two kinds of data passed over the local area work. During the manipulator initialization phase, the modules connect to the work one by one, starting at the base and ending at the endeffector. On joining the work, each module sends a datapacket to the host interface containing its serial number and its relative orientation with respect to the previous module. This information allows us to automatically determine the current manipulator configuration. During the operation phase, the host interface municates with each of the nodes at 400Hz. The data that is exchanged depends on the control mode—centralized or distributed. In centralized control mode, the torques for all the joints are puted on the VMEbased realtime processing unit (RTPU), assembled into a datapacket by the microcontroller on the host interface card and broadcast over the ARMbus to all the nodes of the work. Each node extracts its torque value from the packet and replies by sending a datapacket containing the resolver and tachometer readings. In distributed control mode, on the other hand, the host puter broadcasts the desired joint values and feedforward torques. Locally, in each module, the control loop can then be closed at a frequency much higher than 400Hz. The modules still send sensor readings back to the host interface to be used in the putation of the subsequent feedforward torque. 5 Modular and reconfigurable control software The control software for the RMMS has been developed using the Chimera realtime operating system, which supports reconfigurable and reusable software ponents [15]. The software ponents used to control the RMMS are listed in Table 1. The trjjline, dls, and 8 grav_p ponents require the knowledge of certain configuration dependent parameters of the RMMS, such as the number of degreesoffreedom, the DenavitHartenberg parameters etc. During the initialization phase, the RMMS interface establishes contact with each of the hardware modules to determine automatically which modules are being used and in which order and orientation they have been assembled. For each module, a data file with a parametric model is read. By bining this information for all the modules, kinematic and dynamic models of the entire manipulator are built. After the initialization, the rmms software ponent operates in a distributed control mode in which the microcontrollers of each of the RMMS modules perform PID control locally at 1900Hz. The munication between the modules and the host interface is at 400Hz, which can differ from the cycle frequency of the rmms software ponent. Since we use a triple buffer mechanism [16] for the munication through the dualport RAM on the ARMbus host interface, no synchronization or handshaking is necessary. Because closed form inverse kinematics do not exist for all possible RMMS configurations, we use a damped leastsquares kinematic controller to do the inverse kinematics putation numerically.. 6 Seamless integration of simulation To assist the user in evaluating whether an RMMS con figuration can successfully plete a given task, we have built a simulator. The simulator is based on the TeleGrip robot simulation software from Deneb Inc., and runs on an SGI Crimson which is connected with the realtime processing unit through a Bit3 VMEtoVME adaptor, as is shown in Figure 6. A graphical user interface allows the user to assemble simulated RMMS configurations very much like assembling the real hardware. Completed configurations can be tested and programmed using the TeleGrip functions for robot devices. The configurations can also be interfaced with the Chimera realtime softwarerunning on the same RTPUs used to control the actual hardware. As a result, it is possible to evaluate not only the movements of the man