|Department:||Department of Electrical and Computer Systems Engineering|
|Keywords:||Self-diagnosis; Self-compensation; Autonomous robots; Walking robots|
|Full text PDF:||http://arrow.monash.edu.au/hdl/1959.1/1144154|
This thesis proposes and implements methods for the autonomous identification and classification of disturbances that have negative effects on a robot’s performance (self-diagnosis), and the autonomous selection of suitable compensatory actions (self-compensation). The proposed methods have been implemented in a walking hexapod robot provided with a number of sensors. Both, the robot’s sensorial information and a quantitative measure of the robot’s performance are obtained. This information is used for detecting, identifying and classifying obstructive conditions that have a strong impact on the robot’s performance. Once the cause of a lack of progress in the robot’s mission has been identified, suitable compensatory actions are found, executed and recorded. Then, when previously experienced detrimental situations arise, the associated compensatory measures are immediately taken without involving a searching process. As a result, the recovery from abnormal conditions is accelerated and the robot can promptly continue with its mission. In order to evaluate the performance of the proposed methods, different sets of experiments addressing the robot’s hardware faults, abnormal situations generated in the robot’s environment and a combination of both, were conducted. Results were evaluated by means of two indicators: the number of attempts before a correct identification of the robot’s hardware fault was achieved, and a discrepancy measure. The latter indicates the Euclidean distance between the centroid of an abnormal situation experienced by the robot and the centroid of abnormal situations incorporated into the robot’s database of anomalies. Results showed a good identification rate inside the repertoire of considered abnormal situations. Among the compensatory measures addressed in this thesis, an adaptable gait generation algorithm which allows legged robots to walk in a stable fashion after they have shed a variable number of legs, and a compact leg release mechanism that provides legged robots with a method for the autonomous physical ejection of damaged legs without requiring extra motors, are proposed. Two self-compensating methods (Autonomous Generated Compensatory Measures and Learned Compensating Measures) not implemented into the experimental robot are proposed in this thesis. In theory, these methods will allow robots to autonomously generate compensatory measures and learn from previously encountered abnormal situations. The methods developed in this research are fundamental for the autonomous detection of a robot’s failures and adaptability to unforeseen features of the robot’s environment. By using these techniques, it is expected to increase resilience to damage, extend lifespan and improve autonomy in robotic missions where human intervention is difficult or impossible, such as in extra-terrestrial exploration or other remote hostile environments.