Context-Specific grounding of web natural descriptions to human-centered situations

Human-centered situation, which describes the surrounding world of a person, indicates his undergoing activities. Understanding of human-centered situations helps an assistive robot with its decision making. Existing methods, such as learning from human demonstration, are economically expensive, time-consuming, and have limited scalability. To address this problem, we developed a web-to-situation (W2S) method with which web natural descriptions are grounded into human-centered situations in a context-specific manner. By comparing the learned knowledge from the web and the survey, we proved that W2S is effective in extracting reliable knowledge in an efficient and low-cost manner. By implementing the W2S-retrieved knowledge in 60 web-collected situations and 60 real life situations, we proved that W2S is effective in situation understanding. Given that the web contains huge amounts of information, W2S is expected to effectively scale up a robot's knowledge.