A Simplified Method for Parallelism in Robotics

A robot system consist of robot mechanics and a programmable contrler normaly a digital computer. Such a coamur is serial by nature, ie. it executes one instuction after another, while the robot mechnics are governed by phyical laws which act multaneousy. It is observed that this parllel-serial system requres an iate iput itrface from the paallel robot to the serial computer conuoler (sensed quanies) and outputirfiace back to the robot (actuator commands). This paper suggests an interface with a paralle representation in "I/O memory" of both the input (sensed) q s and output (actuator commands). This I/O nemory is automatically updated in paralel from the sensos iput and loads the outputs into their output representation ultaneou at a predetermined time, if necesary. The update of the I/O memory should accpt asynchronous inputs from both self iniaing input sensors and sensors activated by the control processor. Given this I/O memory representation the robot contro computer can operate intemrly seriay with no tiMing constrat other than it must complete the requred calculations within the allotted time. A robot contoler composed of a set of proceso Ls expected to better match the parallel nature of the robot mechanics than a ingle processor. However, an effcent method has to be used to resolve the couping between the oDmputations on the variu processors. A method to reduce the need of coupling between the comtations of the procesors is to allow the different processor to aces I/O memory in paralel (shared memory). Inho case the shared memory needs to support muliple readers but only sin writers, Le. each I/O memory location has only a single authorized writer. This sipified shared 1/0 memory can be implemented as either a centralized single shared memory or a ditrbuted shared I/O memory. Usually the shired memory is partitioned and the parts distributed between the preors where the interconnect is done by a bus. Now when each processor requires all the data in all the other procesrs' shared memory then the bus traffic is proportional to the number of processors, O(n). To solve this problem it is proposed to distribute the shared memory by providing a copy of the entire I/O memory to each processor and to update all these memories simultaneously by broadcast. Now the traffc on the required broadcast bus is a constant independent of the number of procesors, ie. O(l), and the bus itself is simpler to construct. The design problem of the shared I/O memory raises many questions regarding the way it should be implemented to ensu vabd data, short latency of I/O data transfers and reliable operation of the robot system.