Verifier-Guided Action Selection Improves Embodied Agent Robustness

A team of researchers has introduced VegAS, a framework designed to bolster the resilience of embodied agents utilizing Multimodal Large Language Models (MLLMs) during test time. Rather than opting for a single action, VegAS evaluates a range of potential actions and employs a generative verifier to identify the most dependable choice, all while keeping the original policy intact. The findings indicate that employing a standard MLLM as a verifier does not enhance performance, leading to the development of a data synthesis approach driven by LLMs. This research is available on arXiv with the identifier 2605.12620.