This design-based research project explored how various design features of AR-based learning environments (ARLE) influence students' mathematics self-efficacy and learning of kinematics. Specifically, five ARLEs with different design features were developed and implemented with 136 seventh-grade students in two rounds. Data were gathered from pre- and post-mathematics self-efficacy tests, interviews, classroom observation records and reflective journals, to identify which design features promote or hinder students' learning and self-efficacy. Results revealed that adopting a well-designed ARLE is a viable way to promote students' learning and mathematics self-efficacy, and the success of ARLE lies in (1) coherent tasks with proper time for self-exploration, progressive levels, useful problem contexts and a certain degree of open-endedness, and (2) sufficient technological affordances with strengthened authenticity, clarity and integrated media presentation, as well as engaging interactions with decent flexibility, embodied actions and feedback. These results validated the applicability of the proposed checklist for ARLE designers and practitioners in their design and implementation of ARLEs, within the domain of mathematics, physics and beyond.