Abstract
Expressing confidence is challenging for embodied agents navigating dynamicmultimodal environments, where uncertainty arises from both perception anddecision-making processes. We present the first work investigating embodiedconfidence elicitation in open-ended multimodal environments. We introduceElicitation Policies, which structure confidence assessment across inductive,deductive, and abductive reasoning, along with Execution Policies, whichenhance confidence calibration through scenario reinterpretation, actionsampling, and hypothetical reasoning. Evaluating agents in calibration andfailure prediction tasks within the Minecraft environment, we show thatstructured reasoning approaches, such as Chain-of-Thoughts, improve confidencecalibration. However, our findings also reveal persistent challenges indistinguishing uncertainty, particularly under abductive settings, underscoringthe need for more sophisticated embodied confidence elicitation methods.