Human-robot cooperation is unavoidable in various applications ranging from manufacturing to field robotics owing to the advantages of adaptability and high flexibility. Especially, complex task planning in large, unconstructed, and uncertain environm...
Human-robot cooperation is unavoidable in various applications ranging from manufacturing to field robotics owing to the advantages of adaptability and high flexibility. Especially, complex task planning in large, unconstructed, and uncertain environments can employ the complementary capabilities of human and diverse robots. For a team to be effectives, knowledge regarding team goals and current situation needs to be effectively shared as they affect decision making. In this respect, semantic scene understanding in natural language is one of the most fundamental components for information sharing between humans and heterogeneous robots, as robots can perceive the surrounding environment in a form that both humans and other robots can understand. Moreover, natural-language-based scene understanding can reduce network congestion and improve the reliability of acquired data. Especially, in field robotics, transmission of raw sensor data increases network bandwidth and decreases quality of service. We can resolve this problem by transmitting information in the form of natural language that has encoded semantic representations of environments. In this dissertation, I introduce a human and heterogeneous robot cooperation scheme based on semantic scene understanding. I generate sentences and scene graphs, which is a natural language grounded graph over the detected objects and their relationships, with the graph map generated using a robot mapping algorithm. Subsequently, a framework that can utilize the results for cooperative mission planning of humans and robots is proposed. Experiments were performed to verify the effectiveness of the proposed methods.
This dissertation comprises two parts: graph-based scene understanding and scene understanding based on the cooperation between human and heterogeneous robots. For the former, I introduce a novel natural language processing method using a semantic graph map. Although semantic graph maps have been widely applied to study the perceptual aspects of the environment, such maps do not find extensive application in natural language processing tasks. Several studies have been conducted on the understanding of workspace images in the field of computer vision; in these studies, the sentences were automatically generated, and therefore, multiple scenes have not yet been utilized for sentence generation. A graph-based convolutional neural network, which comprises spectral graph convolution and graph coarsening, and a recurrent neural network are employed to generate sentences attention over graphs. The proposed method outperforms the conventional methods on a publicly available dataset for single scenes and can be utilized for sequential scenes.
Recently, deep learning has demonstrated impressive developments in scene understanding using natural language. However, it has not been extensively applied to high-level processes such as causal reasoning, analogical reasoning, or planning. The symbolic approach that calculates the sequence of appropriate actions by combining the available skills of agents outperforms in reasoning and planning; however, it does not entirely consider semantic knowledge acquisition for human-robot information sharing. An architecture that combines deep learning techniques and symbolic planner for human and heterogeneous robots to achieve a shared goal based on semantic scene understanding is proposed for scene understanding based on human-robot cooperation. In this study, graph-based perception is used for scene understanding. A planning domain definition language (PDDL) planner and JENA-TDB are utilized for mission planning and data acquisition storage, respectively. The effectiveness of the proposed method is verified in two situations: a mission failure, in which the dynamic environment changes, and object detection in a large and unseen environment.