Information Theoretic Intent Disambiguation via Contextual Nudges for Assistive Shared Control

The usefulness of shared-control assistive robots frequently relies on the underlying autonomous agent’s ability to infer human intentions unambiguously, often from low-dimensional and noisy signals generated by the human through a control interface. In this paper, we propose a strategy in which the autonomous agent nudges the context in which the human generates their control actions. In doing so, the autonomous agent attempts to improve its own ability to infer intent accurately, which in turn allows it to provide more accurate assistance. The contributions of this paper are three-fold. First, we introduce an interface-aware information-theoretic metric for active disambiguation that aims to characterize world states according to their potential to extract maximally intent-expressive control actions from the user. Second, we propose a turn-taking based human-autonomy interaction protocol in which the autonomous agent utilizes the disambiguation metric to help itself reduce the uncertainty of its prediction of human intent. Third, we evaluate our metric and interaction protocol both in simulation and with a 9-person human subject study. Our results suggest that disambiguation (a) helps to significantly reduce task effort, as measured by number of mode switches, task completion times, and number of turns executed by the human, and (b) enables the autonomous agent to provide accurate assistance with greater contribution to the overall control signal.