Semantic Likelihood Models for Bayesian Inference in Human-Robot Interaction

Sweet, Nicholas

Graduate Thesis Or Dissertation

Semantic Likelihood Models for Bayesian Inference in Human-Robot Interaction Public Deposited

Analytics

Citations

Citeable URL: https://scholar.colorado.edu/concern/graduate_thesis_or_dissertations/kk91fm055

Abstract

Autonomous systems, particularly unmanned aerial systems (UAS), remain limited in autonomous capabilities largely due to a poor understanding of their environment. Current sensors simply do not match human perceptive capabilities, impeding progress towards full autonomy. Recent work has shown the value of humans as sources of information within a human-robot team; in target applications, communicating human-generated ‘soft data’ to autonomous systems enables higher levels of autonomy through large, efficient information gains. This requires development of a ‘human sensor model’ that allows soft data fusion through Bayesian inference to update the probabilistic belief representations maintained by autonomous systems. Current human sensor models that capture linguistic inputs as semantic information are limited in their ability to generalize likelihood functions for semantic statements: they may be learned from dense data; they do not exploit the contextual information embedded within groundings; and they often limit human input to restrictive and simplistic interfaces. This work provides mechanisms to synthesize human sensor models from constraints based on easily attainable a priori knowledge, develops compression techniques to capture information-dense semantics, and investigates the problem of capturing and fusing semantic information contained within unstructured natural language. A robotic experimental testbed is also developed to validate the above contributions.

Creator