Poor　data-efficiency　and　delayed　reward　limit　the　implementation　of　deep　reinforcement　learning　(DRL)　approach　for　360-degree　video　streaming　in　heterogeneous　wireless　networks.　In　this　paper,　we　leverage　DRL-based　approach　and　priority-based　multi-path　communication　for　360-degree　video　streaming.　First,　we　adopt　soft　actor-critic　for　discrete　action　(SAC-D)　to　decide　network　utilization　ratio　of　each　link,　which　features　a　state-of-art　off-policy　learning　on　the　far　future　decisions　for　multiple　paths.　Second,　we　propose　a　priority-aware　frame　scheduling　to　further　maximize　video　quality.　The　scheduling　can　order　SVC　bit　stream　to　flexibly　utilize　the　spatial　and　quality　characteristics　of　tile-based　approach　and　then　reduce　computational　complexity.　Finally,　we　evaluate　the　proposed　scheme　on　a　semi-physical　test　platform.　Experience　results　show　that　our　algorithm　significantly　outperforms　the　comparison　algorithm　in　terms　of　overall　QoE　quality　and　base　layer　(BL)　freeze　ratio.　©　2021　IEEE.