Abstract
Robots and their control systems are becoming increasingly complex as growing demands are made for their intelligent operation. Automated design processes reduce the complexity involved in designing robots, often leveraging dynamic simulation technology to evaluate potential robot control system designs. However, physics simulators do not provide a perfect representation of the real world. Subsequently, control systems designed in a virtual world will often fail to transfer to the real world. This thesis presents the design, implementation and evaluation of the Physics Abstraction Layer (PAL), a uniform component based software interface to multiple physics engines. PAL can be used to validate the results of an automated design process, increasing the likelihood that a controller will function in the real world. All the physics engines fully supported by PAL were evaluated in a set of benchmarks assessing the key simulation aspects including friction and restitution models, collision detection and response, and the constraint solvers. None of the thirteen physics engines evaluated was found to perform adequately in all aspects. This result indicates that multiple physics engines should be combined when evaluating a controller design to achieve valid results. A genetic algorithm was used to automatically design robot control systems for two application areas. In the first application, a spline controller was evolved for bipedal robot locomotion using the PAL's rigid body simulators and a high fidelity multibody simulator. The controllers evolved using PAL outperformed the controllers evolved using previous approaches. In the second application, a wall following PID control system was evolved for an Autonomous Underwater Vehicle (AUV). The control systems that were evolved using multiple fluid dynamics models outperformed all control ix systems evolved using either a Lagrangian Smoothed Particle Hydrodynamics (SPH) model or a Eulerian model. The biped and
Original language | English |
---|---|
Qualification | Doctor of Philosophy |
Publication status | Unpublished - 2009 |