Abstract: ObjectivesTime-varying propagation and uneven noise lead to severe link asymmetry in underwater acoustic channels. Traditional routing protocols relying on bidirectional link assumptions suffer from neighbor misjudgment, acknowledgment failures and invalid retransmissions. A Q-Learning Based Asymmetric Routing (QBAR) protocol is proposed for underwater sensor networks. Methods QBAR identifies unidirectional and bidirectional neighbors based on link asymmetry characteristics to avoid neighbor detection errors. It constructs a multi-dimensional Q-Learning model incorporating link asymmetry, residual energy and retransmission counts, and designs a dynamic reward function with relay incentives and retransmission penalties. Besides, a cooperative relay acknowledgment mechanism is adopted to solve acknowledgment failures caused by reverse link interruptions. ResultsSimulation results on Aqua-Sim-FG show that QBAR outperforms four typical protocols including QELAR and CARMA, with average packet delivery rate increased by 29.7%, end-to-end delay reduced by 36.8%, and total node energy consumption cut by 17.2%. Conclusions QBAR effectively solves underwater bidirectional link asymmetry, providing a high-robustness routing scheme for large-scale, long-term underwater sensor networks.