Autonomous model building using vision and manipulation

Broun, Alan

Abstract

It is often the case that robotic systems require models, in order to successfully control themselves, and to interact with the world. Models take many forms and include kinematic models to plan motions, dynamics models to understand the interaction of forces, and models of 3D geometry to check for collisions, to name but a few. Traditionally, models are provided to the robotic system by the designers that build the system. However, for long-term autonomy it becomes important for the robot to be able to build and maintain models of itself, and of objects it might encounter. In this thesis, the argument for enabling robotic systems to autonomously build models is advanced and explored. The main contribution of this research is to show how a layered approach can be taken to building models. Thus a robot, starting with a limited amount of information, can autonomously build a number of models, including a kinematic model, which describes the robot’s body, and allows it to plan and perform future movements. Key to the incremental, autonomous approach is the use of exploratory actions. These are actions that the robot can perform in order to gain some more information, either about itself, or about an object with which it is interacting. A method is then presented whereby a robot, after being powered on, can home its joints using just vision, i.e. traditional methods such as absolute encoders, or limit switches are not required. The ability to interact with objects in order to extract information is one of the main advantages that a robotic system has over a purely passive system, when attempting to learn about or build models of objects. In light of this, the next contribution of this research is to look beyond the robot’s body and to present methods with which a robot can autonomously build models of objects in the world around it. The first class of objects examined are flat pack cardboard boxes, a class of articulated objects with a number of interesting properties. It is shown how exploratory actions can be used to build a model of a flat pack cardboard box and to locate any hinges the box may have. Specifically, it is shown how when interacting with an object, a robot can combine haptic feedback from force sensors, with visual feedback from a camera to get more information from an object than would be possible using just a single sensor modality. The final contribution of this research is to present a series of exploratory actions for a robotic text reading system that allow text to be found and read from an object. The text reading system highlights how models of objects can take many forms, from a representation of their physical extents, to the text that is written on them.

Citation

Broun, A. Autonomous model building using vision and manipulation. (Thesis). University of the West of England. Retrieved from https://uwe-repository.worktribe.com/output/910595

• Thesis Type: Thesis
• Publicly Available Date: Jun 7, 2019
• Keywords: robotics, computer vision, active vision
• Public URL: https://uwe-repository.worktribe.com/output/910595
• Additional Information: Additional Information : Supervised by Pipe, T., Mirmehdi, M., Melhuish, C.
• Award Date: Jun 28, 2016