A directed learning-based evolutionary approach for legged robot motion
- Publication Type:
- Issue Date:
At any stage of development, any organism is required to go through a learning phase as a way to acquire and enhance certain skills. Similarly, legged robots perform a learning phase to obtain basic skills before implementing them in real specific applications. Walking skills are one of the first capabilities that need to be learned, as this allows the robots to traverse their environment. This thesis will focus on the learning to walk task for legged robots using machine learning techniques. This task is aimed to develop framed optimal walking gaits which typically concern the speed and stability. In brief, the framed optimal walking gaits are obtained from a particular region in the search space which is framed within the previous robot learning experience or history. We found that the long deployment of the robot at the learning stage which utilises machine learning techniques will eventually raise the significant impact of wear and tear and it is undesirable for robots to be worn out just to perform the learning task before they are deployed in. Typically ML techniques require thousands of trials to find globally optimal solutions, however, this is not an option for robot learn to walk approached since an average robot would be unable to complete the trials intact. In relation to this issue, we propose a Directed Evolutionary Algorithm Learning method - which we henceforth refer to as the DEAL method. This method aims to direct the learning process by incorporating the robot learning experience, and hence, to reduce the time period required for the learning process to converge to a framed optimal solution. This will reduce the deployment time of the robot and as a result, the learning process will have a minimal impact on the wear and tear of the robot body and maximise the life of the robot when it is deployed on specific applications. We empirically conducted literature studies on the four-legged AIBO robot which is a benchmark platform for the international robot soccer competition, RoboCup, and ran the experiment within the same robot platform. The results of the implementation of the DEAL method show that the learning lasts for shorter time than the previous methods used by the University's RoboCup Robot Soccer Team, UTS Unleashed! and as a result, in the long term, the robot productivity will increase through the enhancement of its lifespan and thus its ability to perform other tasks. The objectives of future work will also address the significant factor of wear and tear at any time during robot deployment. The benchmark of the experiment of the DEAL method and associated experimental design will also be expanded to the humanoid robots, the BIOLOID and the NAO robots.
Please use this identifier to cite or link to this item: