Researchers have developed a new method for evaluating the performance of mobile agents, such as robots or drones, that are guided by language. The approach, called CRATE, uses a two-stage framework to assess both task completion and operational safety. It works by analyzing each step of the agent's trajectory and inferring action-conditioned state changes. This information is then aggregated to provide an evidence-grounded evaluation. The method has been shown to be effective in experiments using benchmark datasets. --- Why it matters: This matters because it provides a more nuanced understanding of mobile agent performance, allowing for better task completion and safety assessment. It's particularly relevant for applications where robots or drones are used in complex environments, such as search and rescue missions or environmental monitoring. Source: https://arxiv.org/abs/2608.20797