Browsing Masters e-theses by Subjects
Now showing items 1-4 of 4
The development of test action bank for active robot learningIn the rapidly expanding service robotics research area, interactions between robots and humans become increasingly cornmon as more and more jobs will require cooperation between the robots and their human users. It is important to address cooperation between a robot and its user. ARL is a promising approach which facilitates a robot to develop high-order beliefs by actively performing test actions in order to obtain its user's intention from his responses to the actions. Test actions are crucial to ARL. This study carried out primary research on developing a Test Action Bank (TAB) to provide test actions for ARL. In this study, a verb-based task classifier was developed to extract tasks from user's commands. Taught tasks and their corresponding test actions were proposed and stored in database to establish the TAB. A backward test actions retrieval method was used to locate a task in a task tree and retrieve its test actions from TAB. A simulation environment was set up with a service robot model and a user model to test TAB and demonstrate some test actions. Simulations were also perfonned in this study, the simulation results proved TAB can successfully provide test actions according to different tasks and the proposed service robot model can demonstrate test actions.
Pressure sensor array model for collecting user’s responses to test action in active robot learningActive robot learning (ARL) is an approach to the development of beliefs of the robots on their users’ intentions and preferences, which is needed by the robots to facilitate the seamless cooperation with users. Such approach allows the robots to perform tests on its users and to form high-order beliefs according to the users’ responses. This study carried out primary research on designing a pressure sensor array model attached to the robot’s finger tips to collect the user’s responses to test action in the ARL system. A mathematical model and the reference value threshold which decides the pressure distribution were proposed through a benchmark scenario experiment. The robot holds an object and presents it to the user. When the user does not take over the object, the pressure distribution on the robot’s finger tips shown on the pressure sensor array is uneven. When the user takes over the object, the pressure distribution on the robot’s finger tips is even. According to the relationship between the pressure distribution and the user’s responses, the user’s responses to test action can be recognized by the robot. Two cases of the benchmark scenario which is the robot passing an object to the user is simulated in a simulation software, GraspIt, in this study. The simulation results proved the developed pressure sensor array model can successfully collect the user’s responses to test actions in the ARL.
A statistical approach to a verb vector task classifierHow to enable a service robot to understand its user's intention is a hot topic of research today. Based on its understanding, the robot can coordinate and adjust its behaviours to provide desired assistance and services to the user as a capable partner. Active Robot Learning (ARL) is an approach to the development of the understanding of human intention. The task action bank is part of the ARL which can store task categories. In this approach, a robot actively performs test actions in order to obtain its user's intention from the user's response to the action. This thesis presents an approach to verbs clustering based on the basic action required of the robot, using a statistical method. A parser is established to process a corpus and analyse the probability of the verb feature vector, for example when the user says "bring me a cup of coffee", this means the same as "give me a cup of coffee". This parser could identify similar verbs between "bring" and "give" with the statistical method. Experimental results show the collocation between semantically related verbs, which can be further utilised to establish a test action bank for Active Robot Learning (ARL).
Test moment determination design in active robot learningIn recent years, service robots have been increasingly used in people's daily live. These robots are autonomous or semiautonomous and are able to cooperate with their human users. Active robot learning (ARL) is an approach to the development of beliefs for the robots on their users' intention and preference, which is needed by the robots to facilitate the seamless cooperation with humans. This approach allows a robot to perform tests on its users and to build up the high-order beliefs according to the users' responses. This study carried out primary research on designing the test moment determination component in ARL framework. The test moment determination component is used to decide right moment of taking a test action. In this study, an action plan theory was suggested to synthesis actions into a sequence, that is, an action plan, for a given task. All actions are defined in a special format of precondition, action, post-condition and testing time. Forward chaining reasoning was introduced to establish connection between the actions and to synthesis individual actions into an action plan, corresponding to the given task. A simulation environment was set up where a human user and a service robot were modelled using MATLAB. Fuzzy control was employed for controlling the robot to carry out the cooperative action. In order to examine the effect of test moment determination component, simulations were performed to execute a scenario where a robot passes on an object to a human user. The simulation results show that an action plan can be formed according to provided conditions and executed by simulated models properly. Test actions were taken at the moment determined by the test moment determination component to find the human user's intention.