NRI: FND: Knowledge-based Robot Sequential Decision Making under Uncertainty

Suny At Binghamton, Binghamton NY

Investigators

Abstract

This project focuses on promoting the progress of science and technology development by developing computational frameworks needed to advance research on robot decision-making. The key activity is building intelligent agents that are able to perceive the environment through sensors and act upon that environment through actuators. This project aims to bring in computational methods of different modalities toward a generally applicable, robot decision-making framework that can significantly promote the development of intelligent agents in the real world. Example application domains include robotics (as mostly used in this proposal), finance, urban planning, healthcare, games, transportation, e-commerce, and many more. A key focus of this project is on incorporating robotic projects into K-16 education, including education of undergraduate students through the University Undergraduate Research (UUR) programs, which aim at exposing scientific research experiences to undergraduate students in early years, outreach education to regional high/middle schools, and education to the general public through public media. Robust robot decision-making in the real world is challenging for reasons such as imperfect perception capabilities, incomplete domain knowledge, non-deterministic action outcomes, and limited experience of interacting with the working environment. The focus of this proposed project is on a robot sequential decision-making framework that simultaneously supports learning to perceive the environment, reasoning about declarative contextual knowledge, and planning to actively collect information for task completion. Under the umbrella of artificial intelligence, there are at least three very different ways of realizing robot decision-making, namely supervised learning from robot experiences in the past, automated reasoning using declarative knowledge, and probabilistic planning toward accomplishing complex tasks that require more than one action. The research aims include developing algorithms that simultaneously allow logical-probabilistic reasoning and planning under uncertainty for robot decision-making in stochastic worlds; developing algorithms for simultaneous world state estimation via recurrent neural networks, representing and reasoning with declarative knowledge, and multi-modal perception control using decision-theoretic methods; and developing a principled integration of logical-probabilistic knowledge representation and reasoning (KRR) with reinforcement learning, enabling agents to simultaneously reason with declarative knowledge and learn from interaction experiences. This work has the potential to enable visionary robot decision-making while leveraging the extensive interaction experiences as well as contextual knowledge from people. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.