Resumen de Locomotion through morphology, evolution and learning for legged and limbless robots
Robot locomotion is concerned with providing autonomous locomotion capabilities to mobile robots. Most current day robots feature some form of locomotion for navigating in their environment. Modalities of robot locomotion includes: (i) aerial locomotion, (ii) terrestrial locomotion, and (iii) aquatic locomotion (on or under water). Three main forms of terrestrial locomotion are, legged locomotion, limbless locomotion and wheel-based locomotion. A Modular Robot (MR), on the other hand, is a robotic system composed of several independent unit modules, where, each module is a robot by itself. The objective in this thesis is to develop legged locomotion in a humanoid robot, as well as, limbless locomotion in two-dimensional (2D) modular robotic configurations. Taking inspiration from biology, robot locomotion from the perspective of robot morphology, through evolution, and through learning are investigated in this thesis.
Locomotion is one of the key distinguishing characteristics of a zoological organisms. Almost all animal species, and even some plant species, produce some form of locomotion. In the past few years, robots have been “moving out” of the factory floor and research labs, and are becoming increasingly common in everyday life. So, providing stable and agile locomotion capabilities for robots to navigate a wide range of environments becomes pivotal. Developing locomotion in robots, through biologically inspired methods, furthers our understanding on how biological processes may function.
Connected modules in a configuration, exert force on each other, as a result of interaction between each other, and their environment. This phenomena is studied and quantified, and then used as implicit communication between robot modules for producing locomotion coordination in MRs. Through this, a strong link is established between the morphology of the robot, and the gait that emerge in it.
A variety of locomotion controller, both periodic function based and morphology based, are developed for MR locomotion and bipedal gait generation. A hybrid Evolutionary Algorithm (EA) is implemented for evolving gaits, both in simulation as well as in the real-world, on a physical robot. Limbless gaits in MRs are also learnt by learning optimal control policy, through Reinforcement Learning (RL) method.
