Egocentric Action Recognition (EAR) has gained significant attention due to its widespread applicability in lifestyle analysis, medical monitoring, and industrial robotics, among other real-world scenarios. However, existing EAR methods are built on the closed-set assumption, making it challenging to handle unknown actions inevitably present in open-world scenarios and struggling to meet the dual requirements of accuracy and reliability while providing decisions. To address the Open-set EAR problem, this paper presents a Open-set Egocentric Action Recognition (OpenEAR) framework, advancing beyond traditional egocentric action recognition methods. OpenEAR distinguishes itself by adeptly handling unknown actions in open-world scenarios, a notable limitation in conventional EAR models. Utilizing large-scale pre-trained models and refined architecture, OpenEAR excels in semantic extraction from egocentric videos. Its unique incorporation of Evidential Deep Learning (EDL) allows for uncertainty estimation, enhancing prediction reliability. This novel approach not only recognizes known actions and objects but also quantifies prediction confidence, effectively managing unknown elements. Demonstrated superior performance on EPIC-KITCHENS-55 and EGTEA Gaze+ datasets underlines OpenEAR’s robustness and practicality, marking a significant leap from existing methods. The OpenEAR framework is available at https://github.com/zou-y23/OpenEAR .