- Title
- Static and bootstrapped neuro-simulation for complex robots in evolutionary robotics
- Creator
- Woodford, Grant Warren
- Subject
- Robotics
- Date Issued
- 2019
- Date
- 2019
- Type
- Thesis
- Type
- Doctoral
- Type
- PhD
- Identifier
- http://hdl.handle.net/10948/44656
- Identifier
- vital:38172
- Description
- Evolutionary Robotics (ER) is a field of study focused on the automatic development of controllers and robot morphologies. Evolving controllers on real-world hardware is time-consuming and can damage hardware through wear. Robotic simulators can be used as an alternative to a real-world robot in order to speed up the ER process. Most simulation techniques in practice use physics-based models that rely on an understanding of the robotic system in question. Developing effective physics-based simulators is time consuming and requires a significant level of specialised knowledge. A lengthy simulator development and tuning process is typically required before the ER process can begin. Artificial Neural Networks simulators (SNNs) can be used as an alternative to a physics based simulation approach. SNNs are simple to construct, do not require significant levels of prior knowledge of the robotic system, are computationally efficient and can be highly accurate. Two types of ER approaches utilising SNNs exist. The Static Neuro-Simulation (SNS) approach involves developing SNNs before the ER process where these SNNs are used instead of a physics-based simulator. Alternatively, SNNs can be developed during the ER process, called the Bootstrapped Neuro-Simulation (BNS) approach. Prior work investigating SNNs has largely been limited to simple robots. A complex robot has many degrees of freedom and ifa low-level controller design is used, the solution search space is high-dimensional and difficult to traverse. Prior work investigating the SNS and BNS approaches have mostly relied on simplified controller designs which rely on built-in prior knowledge of intended robot behaviours. This research uses low-level controller designs which in turn rely on low level simulators. Most ER studies are conducted on a single type of robot morphology. This research investigates the SNS and BNS approaches on two significantly different classes of robots. A Hexapod and Snake robot are used to study the SNS and BNS approaches. The Hexapod robot exhibits limbed, walking behaviours. The Snake robot is limbless and generates crawling behaviours. Demonstrating the viability of the SNS and BNS approaches for two different classes of robots provides strong evidence that the tested approaches are likely viable on other classes of robots. Various proposed improvements to the SNS and BNS approaches are investigated. The Results demonstrate that the SNS and BNS approaches are viable when applied to Hexapod and Snake robots without restricting controller designs to those with significant levels of built-in prior knowledge of robot behaviours. SNNs configured in ensembles improve the likely performance outcomes of solutions. The expected benefit of adding simulator noise during the evolutionary process were not as pronounced for problems investigated in this work.
- Format
- xiii, 306 leaves
- Format
- Publisher
- Nelson Mandela University
- Publisher
- Faculty of Science
- Language
- English
- Rights
- Nelson Mandela University
- Hits: 825
- Visitors: 737
- Downloads: 28
Collections
NMU School of Computer Science, Mathematics, Physics and Statistics
| Thumbnail | File | Description | Size | Format | |||
|---|---|---|---|---|---|---|---|
| View Details Download | SOURCE1 | Woodford, G 205104224 Thesis Dec 2019.pdf | 18 MB | Adobe Acrobat PDF | View Details Download |