Robotics education has gradually been emphasized in contemporary school curricula; however, assessment tools for robotics learning are still limited. Based on Bloom's Taxonomy of educational objectives, this study aimed to develop the Robotics Learning Self-Efficacy Scale (RLSES) with a two-level construct of five dimensions for assessing students' self-efficacy for learning robotics. A total of 181 elementary, junior high and senior high school students (5th-12th graders) with robotics learning experience were selected as the sample of this study. A questionnaire including 32 candidate items designed for the initial version of the RLSES was administered to the sample. An exploratory factor analysis was conducted and, finally, 16 items were drawn for the final RLSES under five subscales ("Comprehension," "Practice," "Analysis," "Application," and "Collaboration"), with a total explained variance of 85.28%. The Cronbach's alpha reliability was 0.97 for the overall scale, ranging from 0.87 to 0.95 for the subscales. The inter-correlation analysis showed evidence of discriminant validity. Regression analysis results supported that "Practice" and "Comprehension" self-efficacy were significant predictors of "Analysis," "Application," and "Collaboration" self-efficacy, confirming the two-level (2 × 3) construct of the RLSES. Significant differences among school levels were found and are discussed.