Monocular　visual-inertial　state　estimator　can　be　classified　into　filter-based　and　optimization-based　method.　However,　both　of　these　methods　require　an　initial　state　to　bootstrap　the　system.　In　this　paper,　we　propose　a　robust　initialization　algorithm　to　provide　high-quality　initial　guess　for　the　monocular　visual-inertial　system　(VINS).　The　proposed　method　takes　the　up-to-scale　camera　poses　estimated　by　monocular　vision-only　SLAM　(simultaneous　localization　and　mapping)　and　the　IMU　sensor　measuements　as　input.　The　gyroscope　bias　are　firstly　estimated　by　minimizing　the　error　between　pre-integrated　gyroscope　measurements　and　camera　attitude　measurements.　After　that,　a　rough　gravity　vector　is　calculated　by　aligning　the　pre-integrated　measurements　and　camera　translation　measurements.　The　observable　of　the　accelerometer　bias　is　also　checked　in　this　step.　The　gravity　vector　is　refined　in　the　third　step,　with　the　velocity　is　estimated　in　this　step　as　well　as　the　acceleration　bias　if　it　is　observable.　The　practicability　of　the　proposed　approach　is　validated　by　simulation　and　real　datasets　experiments.　©　2018　Association　for　Computing　Machinery.