A conceptual framework for hallucination detection and mitigation in VLM-driven embodied agents
Abstract
Embodied agents driven by vision-language models (VLMs) can describe a scene fluently and invent objects, attributes, relations or executable targets. Such errors may cause serious injury if the language response is translated into navigation or manipulation operations. This paper introduces EViD-M as a conceptual computer-vision framework based on evidence-aware VLMs and embodied-perception research. Organise atomic-claim scoring via region alignment, RGB-D geometry, temporal memory, counterfactual masking and token uncertainty. A risk-aware decoder is used to modify unsupported claims, issue abstentions, and forbid unsafe behaviour until a confirming view is obtained. The framework specifies a scene-isolated EHA-Indoor stress-set protocol with 1,200 photorealistic RGB-D episodes covering object absence, attribute swaps, relation swaps and occlusion. AUROC, expected calibration error, hallucination rate, unsafeaction rate and task success are defined as evaluation indicators, and the numerical profiles presented below serve as reference values for subsequent empirical verification. Conduct physical-robot experiments to evaluate the application effect.