Artificial neural networks for learning inverse kinematics of humanoid robot arms

Official URL: http://risc01.sabanciuniv.edu/record=b1615008 (Table of Contents)

Nowadays, many humanoid teen sized robot platforms have been developed by di erent research groups. The idea is either to conduct research or to produce a speci c task ful lling machine. This imposes many challenges on the design of algorithms for di erent actions like walk or reaching some targets. There are many sophisticated humanoid research platforms available, but one crucial aspect to look is the developmental cost associated with the task. As the name describes, the Humanoid robots are the ones that resemble humans in their design as well as their way of performing the task. In the development of humanoid robots, many design for the arm of a humanoid robot has been studied. We have developed an arm with 5 degrees of freedom using dynamixel servo motors. We used 3D plastic printing for manufacturing the part. This arm with multiple degrees of freedom enables the robot to have free movement around the body. Besides, we also designed a simulator model of a robot that works with the advanced simulators available today. A great number of approaches and algorithms have been implemented to solve the problem of inverse kinematics. The research carried out in this thesis takes the early learning in human infants as the basis. Human infants in their early age of development move their arm to reach new goals that they have not seen yet and with the help of the visual feedback they learn the limits and possibilities of reaching targets. We have used this idea to develop a learning algorithm that eventually enables the robot to reach goals in 3D space accurately. This algorithm is advantageous in the sense that it is faster than the parent approach of Rolf [2013] and no prior knowledge of the arm model is required to learn the inverse solution for correct positioning. The algorithm starts with the knowledge of only one goal in the 3D space, explores more goals in the 3D space and the learning enables the algorithm to grasp the solution of inverse positioning of the arm. The results obtained are comparable to the results generated by Rolf [2013] with the advantage that the learning is fast with our algorithm. In current research in the eld of cognitive and developmental robotics, one aim is to develop robots based on biological beings (for example humans and animals) present on our planet. Humanoid robots can be considered as an exemplary development in this sense. Similarly, the researchers are trying to move the mathematically computational solutions more towards bio-inspired computational solutions. Therefore, exploring bio-inspired learning which was achieved by taking advantage of Arti cial Neural Networks (ANNs) is another advantage associated with this work.