AbstractDemand for robotic solutions to problems in every area of society has been rapidly increasing. Some notable examples include providing independence and care to the elderly, managing degrading farm land, and reducing the risk to human life with agile robots for search and rescue. Many of these areas will require robots to either work alongside humans, or in environments that are unstructured. For these an artificial sense of touch will be crucial.
Tactile sensing is a young field of research when compared to other sensing fields like computer vision. But just like computer vision, tactile sensing will open up important doors for robotics systems. Tactile provides an active sense for robots, allowing contact level perception on its influence on the world. It is this sensing and understanding of environmental influence that can be utilised for manipulation.
This thesis presents novel developments in tactile sensing hardware, perception, and deployment. The work demonstrates the development of the TacTip tactile sensor to be better suited to rapid prototyping and complex morphologies. This is achieved by redeveloping the TacTip technology to be bot modular and 3D printable in a multi-material printer. Ultimately this allowed the for exploration the effects of biomimetic fingerprints on tactile perception of varies spatial scales. Demonstrating improvements in acuity of location perception with its inclusion.
An investigation into using active perception algorithms for active manipulation is explored. Where the principle that existing algorithms that provide control for perception can be used such that perception for control is achieved. The work demonstrated the successful rolling of a cylinder on a table top using only tactile sensing, and highlights that the methods have a trade off between accuracy and reaction time.
To improve the generality of the TacTip sensors tactile sensing, I present the development of a novel method for inferring a third dimension to the sensor data. This method deploys the mathematical principle of voronoi tessellation to the point data outputted from the sensor. This tessellation creates cells around each point, the areas of which can be interpolated to crate a 3D surface representation of the data. Ultimately, along with exploration of the raw point data, tactile features such as shear, pressure, and contact locations could be inferred with out the use of data intensive machine learning techniques.
Lastly, this thesis present a fully tactile seven degree of freedom hand, fully equipped with the new TacTip developments and generalised feature inference. The hand was designed to be highly tactile, dexterous and relatively inexpensive tool. The hand is benchmarked on the YCB objectset with a closed loop adaptive grasp controller which demonstrates its viability for starting to explore tactile dexterous manipulation.
Overall this thesis demonstrates developments in tactile sensing with accurate location perception, feature perception, simple manipulations, and grasp adaptation. All of these are components necessary for reaching the ultimate goal and bigger challenge of complex dexterous tactile manipulation.
|Date of Award
|1 Oct 2019
|Nathan F Lepora (Supervisor) & Arthur G Richards (Supervisor)
- Robot Hand