Semantic based task planning for domestic service robots

Task Planning is developed for an autonomous mobile robot in order to support the robot to accomplish tasks in various degrees of environmental complexity. This environment can be fixed or deterministic (as in a factory), dynamic (as in the human domestic household), or non-deterministic (as in the space exploration). The robot should be provided with a reliable planning system in order to face its major challenge of being certain that its plan to accomplish a task is generated correctly, regardless of the dynamic or uncertain elements of its environment. This thesis is focused on providing the robot task planner with the ability to generate its plans reliably and detect the failures in generating correct plans. Previous approaches for generating plans depended mainly on action effects (explicit effects) that are encoded in the action model. This means that the action effects should cover most of the characteristics of the newly generated world state. However, this extra information can complicate the action model, especially in the real world. In this thesis, a semantic knowledge base is proposed to derive and check implicit information about the effects of actions during plan generation. For example, this approach would inform the robot, that it had entered a bedroom because it has recorded at least one bed and zero ovens. When a robot enters a room, the implicit expectations are derived from a semantic knowledge base about that type of room. These expectations should be verified in order to make sure the robot is in the correct room. The main contributions of this thesis are as follows:  The concept of using the Semantic Knowledge Base (SKB) to support the robot task planner under deterministic conditions has been defined. A new model of high-level robot actions has been developed, and this model represents the details of robot action as ontology. This model is thus known as the Semantic Action Model (SAM). An algorithm that integrates SKB and SAMs has also been developed. This algorithm creates the “planning domain” in the Planning Domain Definition Language (PDDL) style. This is used as input to the planner to generate the plan for robot tasks. Then, a general purpose planning algorithm has also been defined, which can support planning under deterministic conditions, and is based on using ontology to represent SKB. ii  A further contribution relates to the development of a probabilistic approach to deal with uncertainty in semantic knowledge based task planning. This approach shows how uncertainties in action effects and world states are taken into account by the planning system. This contribution also served to resolve situations of confusion in finding an object relevant to the successful generation of an action during task planning. The accuracy related to this type of planning in navigation scenario, on average, is (90.10%).  An additional contribution is using the planning system to respond to unexpected situations which are caused by lack of information. This contribution is formalised as a general approach that models cases of incomplete information as a planning problem. This approach includes a sequence of steps for modelling and generating a plan of actions to collect the necessary information from the knowledge base to support the robot planner in generating its plan. This results in developing a new type of action which is known as a Semantic Action Model for Information Gathering (SAM_IG). These actions have the ability to access the knowledge base to retrieve the necessary information to support the planning system when it is faced with incomplete information. The information gathering approach is also used to gather the necessary information in order to check the implicit expectations of the generated actions. The correct classification related to this type of planning in navigation scenario, on average, is (92.83 %).  Another contribution is concerned with solving the problem of missing information, which is using the methods for measuring concept similarity in order to extend the robot world state with new similar objects to the original one in the action model. This results in developing Semantic Realisation and Refreshment Module (SRRM) which has the ability to estimate the similarities between objects and the quality of the alternative plans. The quality of the alternative plans could be similar to the original plan, in average, 92.1%. The results reported in this thesis have been tested and verified in simulation experiments under the Robot Operating System (ROS) middleware. The performance of the planning system has been evaluated by using the planning time and other known metrics. These results show that using semantic knowledge can lead to high performance and reliability in generating robot plans during its operation