Learning Tasks and Missions from Natural Language Instructions in One Shot

M. Scheutz
Thinking Robots, Inc.,
United States

Keywords: One-shot learning, natural language understanding, cognitive robotic architecture

Summary:

Current autonomous systems are inflexible, do not understand human objectives, and cannot quickly learn new knowledge and capabilities needed in open-world missions. Our software enables flexible, instructible systems that can quickly learn new knowledge about objects, actions, relations, events, rules, goals, etc. through natural language dialogues without the need to collect large amounts of data or use any computationally expensive methods for training (as is necessary with deep neural networks). The software architecture is fully introspectible, hence the system can report what is does and does not know, can be modified by authorized users online, can detect faults and attempt to mitigate them, generate performance estimates for actual and hypothetical scenarios, and produce explanations for its decisions and behaviors. The software is generic and can be used with any number of heterogeneous autonomous systems allowing for natural human-machine teaming.