Relearning robot control : discovering muscle pairs of a humanlike robotic arm

Research in humanlike robotic arms will not only provide insight in how to make more realistic looking and behaving robots, but can also provide insight into the way that humans control their muscles. This information can then be applied in rehabilitation projects or in the development of prosthetics. Most current robots have joint-based actuators, instead of more humanlike systems with artificial muscles. Controlling humanlike systems requires a different approach than those used for conventional robots because the muscles of humanlike muscle-based systems have overlapping effects, while actuators conventional joint-based systems only affect one joint angle. This redundancy in control is of interest for adaptability to new environments and to changes in the body like wear or other damage. This thesis presents a simple method to control a humanlike robotic arm that does not require a lot of information about the arm itself or about its environment. The goal of this thesis is to find pairs of muscles that can have an opposite effect for a specific task. These pairs are useful for higher level control, such as compliance. Results show that the wrist of the arm can be positioned using only the learned muscle pairs, even though the pairing is not always optimal. Furthermore, the inaccuracies due to approximations can be mitigated by using the Attractor Selection Model to handle noise in the robot arm and the environment.