With　its　unique　intrinsic　flexibility　and　muscle-like　output　force　characteristics,　Pneumatic　Artificial　Muscle　(PAM)　has　gained　important　applications　in　the　driving　of　Coexisting-Cooperative-Cognitive　Robots　(Tri-Co　Robots).　However,　due　to　the　difficulty　of　accurately　modeling　caused　by　the　severe　non-linear　characteristic　of　PAM,　model-based　control　methods　are　not　easy　to　achieve　ideal　control　results.　Therefore,　this　paper　proposes　a　Model-free　Reinforcement　Learning　(RL)　method　to　control　a　bionic　single　joint　actuated　by　antagonistic　PAMs.　We　designed　the　angle　tracking,　angle　positioning　and　variable　stiffness　experiments　of　the　robot　joint,　and　achieved　satisfactory　human-like　motion　effects.　The　research　in　this　paper　can　effectively　improve　the　control　performance　of　robotic　joint　actuated　by　antagonistic　PAMs,　thereby　improving　the　accuracy,　compliance　and　safety　of　the　motion　control　of　Tri-Co　Robots.