Deploying language-driven agents in robotics requires evaluations that reflect real-world task demands: natural-language instructions with reproducible outcomes. Such agents must connect language models to callable perception and control tools, and be assessed using deployment-critical metrics including latency, accuracy, and error modes. We present SCOPE (Simulation and Camera Operations for Perception and Evaluation), a modular agent for natural-language, open-vocabulary pan–tilt–zoom (PTZ) camera control and visual scene understanding, designed explicitly for edge deployment. SCOPE operates both in a Blender-based simulation environment and on a physical PTZ camera, executing all perception, planning, and control locally at the deployment site using edge-accessible compute.

Low Earth Orbit (LEO) satellite constellations, led by deployments such as Starlink, are playing an increasingly pivotal role in enabling global broadband connectivity. However, the reliability and performance of these space-based networks are highly sensitive to environmental dynamics, particularly localized weather phenomena that exhibit strong spatio-temporal variability. In this study, we present a continental-scale geospatial analysis of weather-induced performance degradation in the Starlink LEO network, with a focus on the contiguous United States.

Textual Question Answering (TQA) remains a formidable challenge, despite over a decade of research. The integration of transformer networks and external knowledge via pre-trained models has marked a significant advancement in TQA. Yet, a crucial element often overlooked is the incorporation of external visual understanding. In this study, we introduce an innovative TQA approach that equips machines with the capability for on-demand visual grounding, thereby enriching their comprehension of questions and enhancing the relevance of generated answers. Our methodology utilizes web image search to tap into a vast pool of global knowledge and employs a novel technique for determining the most appropriate answer through on-demand visual grounding. We present a variety of multimedia model configurations, showcasing that our proposed method not only surpasses existing systems without necessitating pre-training but also achieves performance comparable to fine-tuned models 30 times its size as well as closed-source LLMs such as GPT-4o, a testament to its efficiency. Furthermore, an interpretability analysis reveals the integral role of visual grounding in the model’s decision-making process. This research offers a fresh outlook on augmenting TQA performance by harnessing the potential of visual grounding, with broad implications for natural language processing and artificial intelligence.