Reflex based walking pattern adaptation for biped robots

Official URL: http://risc01.sabanciuniv.edu/record=b1081779 (Table of Contents)

Employing robots to replace humans in heavy and dangerous tasks is an important research area. Biped robots have advantages in obstacle avoidance and are therefore suitable to work in the human environment in such tasks. However, their control is a very difficult problem because of their nonlinear and unstable nature. Even very small disturbances can lead to instability. Disturbances can vary from slippery ground surfaces to collisions and unexpected contact with the environment to variations in the payload. For dynamically stable robots (walking on two or less feet), constraints on timing and foot placement increase the difficulty of designing controllers that can anticipate changes in the payload or react to errors. This thesis demonstrates the effectiveness of preprogrammed high-level responses to locomotion in a complex dynamic environment. A suite of responses allows a simulated, three dimensional, bipedal robot to recover from falling down due to a sudden change in the payload. Many environment contact errors would be avoided if the control system can respond fast to the errors that have already taken place and adapt the biped locomotion. In the case of the biped robot considered in this work, the controller might have less than a few tenths of a second in which to choose or plan an appropriate recovery. In this thesis reflexes are defined as responses with no explicit modeling and limited sensing. That is the robot can detect the payload change and makes no attempt to estimate the properties of the load to calculate a corresponding recovery plan. These reflexes are defined at high level because they involve changes of the biped body configuration and trajectory. Sensing elements are used just to detect the error and trigger the reflex. Explicit dynamic modeling of the biped robot is complicated and the controller cannot use it to compute precise and appropriate reactions. In addition, accurate and precise information on load addition is not available to the controller. The method presented changes the walk trajectory and shifts the center of gravity to keep the balance of the walk. Thereafter, the original trajectory is brought back by a smooth trajectory interpolation function. The reflex-adaptation technique considered is tested for a variety of payloads at different loading times. The method shows a good functionality by recovering the biped and allowing stable and balanced original walking pattern. The approach is successful and is a candidate for real applications.