The Universal Manipulation Interface (UMI) is a data collection and strategy learning framework that enables direct transfer of skills from human demonstrations on-site to deployable robot strategies. UMI combines handheld grips with meticulous interface design to achieve portable, low-cost, and informative data collection, used for challenging two-handed and dynamic operation demonstrations. To promote deployable strategy learning, UMI incorporates well-designed strategy interfaces with real-time delay matching and relative trajectory action representation. Thus, the learning strategies are hardware-agnostic and can be deployed on multiple robot platforms. Equipped with these features, the UMI framework opens up new capabilities for robot manipulation, allowing for generalized dynamic, two-handed, precise, and long-duration behaviors, achievable with zero adjustments. We have demonstrated the versatility and effectiveness of UMI through comprehensive real-world experiments, where UMI strategies trained with various human demonstrations achieved zero-shot generalization in new environments and with new objects.