Orion: A Unified Visual Agent (arxiv.org)

Orion is a new "visual agent" that moves beyond descriptive vision-language outputs by orchestrating a toolbox of specialized computer-vision modules to perform precise, multi-step visual reasoning and execution across images, video, and documents. Instead of producing freeform captions, Orion coordinates detectors, keypoint localization, panoptic segmentation, OCR and geometric-analysis routines as callable tools, enabling it to carry out complex workflows (e.g., extract text, locate parts, measure geometry and act on results) in an agentic loop that combines neural perception with symbolic execution. This tool-augmented architecture yields production-grade visual intelligence: the paper reports competitive results on multimodal benchmarks such as MMMU, MMBench, DocVQA and MMLongBench, and showcases generalization to long-form and multi-frame tasks. For the AI/ML community, Orion signals a practical shift from monolithic VLMs toward modular, planner+tool systems that offer greater precision, composability and interpretability for real-world applications (document understanding, video analysis, robotics). The release includes demos, code and examples, making it a concrete reference design for building autonomous, tool-driven visual agents.