"In recent years, a new generation of cognitive robots start to enter our lives. Robots such like ASIMO, PR2, and Baxter have been studied and applied in education and service applications. Different from traditional industry robots doing specific repetitive tasks in a well controlled environment, cognitive robots must be able to work with human partners in a dynamic environment which is filled with uncertainties and exceptions. It is unlikely to pre-program every type of knowledge (e.g., perceptual knowledge like different colors or shapes; action knowledge like how to complete a task) into the robot systems ahead of time. Just like how children learn from their parents, it's desirable for robots to continuously acquire knowledge and learn from human partners on how to handle novel and unknown situations. Driven by this motivation, the goal of this dissertation is to develop approaches that allow robots to acquire and refine knowledge, particularly, knowledge related to verbs and actions, through interaction/dialogue with its human partner. Towards this goal, this dissertation has made following contributions i . As a first step, we propose a goal state based verb semantics and develop a three-tier action/task knowledge representation. This representation on one hand supports the connection between symbolic representations of language and continuous sensori-motor representations of the robots; and on the other hand, supports the application of existing planning algorithms to address novel situations. Our empirical results have shown that, given this representation, the robot can immediately apply the newly learned action knowledge to perform actions under novel situations. Secondly, the goal state representation and the three-tier structure are integrated into a dialogue system on board of a SCHUNK robotic arm to learn new actions through human-robot dialogue in a simplified blocks world. For a novel complex action, the human can give an illustration through dialogue using robot's existing action knowledge. Comparing the environment changes before and after the action illustration, the robot can identify a goal state to represent the novel action, which can be immediately applied to new environments. Empirical studies have shown that action knowledge can be acquired by following human instructions. Furthermore, the results also demonstrate that step-by-step instructions lead to better learning performance compared to one-shot instructions. To solve the insufficiency issue of applying the single goal state representation in more complex domains (e.g., kitchen and living room), the single goal state is extended to a hierarchical hypothesis space to capture different possible outcomes of a verb action. Our empirical results demonstrate that the representation of hypothesis space, combined with the learned hypothesis selection algorithm, outperforms approaches using single hypothesis representation. Lastly, we address uncertainties in the environment for verb acquisition. Previous works rely on perfect environment sensing and human language understanding, which does not hold in real world situation. In addition, rich interactions between teachers and learners as observed in human teaching/learning have not been explored. To address these limitations, the last part presents a new interactive learning approach that allows robots to proactively engage in interaction with human partners by asking good questions to handle uncertainties of the environment. Reinforcement learning is applied for the robot to acquire an optimal policy for its question-asking behaviors by maximizing the long-term reward. Empirical results have shown that the interactive learning approach leads to more reliable models for grounded verb semantics, especially in the noisy environments."--Pages ii-iii.
Read
