Research On Service Robot Neural Network Force Controller Realization

The continuous development of robot requires that robot can have response just like a human being in contacts. This thesis investigates the problem of the realization of a neural network force controller for a service robot, which aims at a human-like service robot.The necessity of robot force control is presented. Several realization methods are compared. Among them, neural network method is more suitable for the requirement of service robot control assignment.A humanoid-like response service robot system is highly non-linear and its parameter varies with time. Its assignment is complex and its environment is unknown. Using neural network for robot force control has many advantages, and need not build a model of the system, and can work when the system is unknown. And neural network has excellent tolerance. Implementing force control for a service robot calls for high real-time characteristic. Using hardware neural network to implement control can achieve better performance.This thesis realizes the computation of sigmoid function, which is a commonly used non-linear transfer function in BP network, using a STAM (Symmetric Table and Addition Method) method. Programmable technology develops rapidly, and is becoming the first selected EDA design method. Here using Xilinx FPGA device implements the arithmetic. At present, there are several methods for non-linear function realization. But these algorithms commonly need to engross larger space, or have long reaction time. So they are unsuitable for hardware realization. However, STAM method significantly reduces hardware resource taken up by the lookup table, and need not to execute iteratively. And it can rapidly fulfill one time computation in an instruction period. The computation error is less than one ulp. Compared to other algorithms, STAM method is more suitable for hardware realization.In addition, many other problems also exist in hardware neural network, including error problem, learning mode, parallel architecture, and also neural network inner linking problem, hidden layer and the realization of the multiplicatorand etc. For instance, error problem: hardware neural network employs the limited precision, and will inevitably bring limited precision error. Choosing appropriate precision can meet the space requirement and not affect the implementation of the network. These issues all should be carefully considered before establishing a network. All above have been investigated in the paper, and the architecture will be taken in the design based on target system.At the end of the paper, based on an implemented robot motion control system, a fabrication, realizing a humanoid-like response service robot, is presented.