A study on automatic gait parameter tuning for biped walking robots

Official URL: http://risc01.sabanciuniv.edu/record=b1081780 (Table of Contents)

Automatic gait parameter tuning for biped walking robots is the subject of this thesis. The biped structure is one of the most versatile ones for the employment of mobile robots in the human environment. Their control is challenging because of their many DOFs and nonlinearities in their dynamics. Open loop walking with offline walk pattern generation is one of the methods for walking control. in this method the reference positions of the foot centers with respect to the body center are generated as functionals. Commonly, the tuning process for the trajectory generation is based on numerous trial and error steps. Obviously, this is a time consuming and elaborate process. In this work, online adaptation schemes for one of the trajectory parameters, "z-reference asymmetry", which is used for the compensation of uneven weight distribution of the robot in the sagittal plane, is proposed. In one of the approaches presented, this parameter is tuned online. As an alternative to parameter tuning, a functional learning scheme employing fuzzy identifiers is tested too. Fuzzy identifiers are universal function approximators. Fuzzy system parameters are adapted via back-propagation. An on-line tuning scheme for biped walk parameters however can only be successful if there is sufficient time for training without falling. The training might last hundreds of reference cycles. This implies that a mechanism for keeping the robot in continuous walk, even when the parameter settings are totally wrong, is necessary during training. In this work, virtual torsional springs which resist against deviations of the robot trunk angles from zero, are attached to the trunk center of the biped. The torques generated by the springs serve as the criteria for the tuning and help in maintaining a stable and a longer walk. The springs are removed after training. This novel approach can be applied to a wide range of control systems that involve parameter tuning. 3-D simulation techniques using C++ are employed for the model of a 12-DOF biped robot to test the proposed adaptive method. in order to visualize the walking, simulation results are animated using an OpenGL based animation environment. As a result of the simulations, a functional for the desired parameter, keeping the system in balance while walking, is generated.