Abstract: To address the challenges of complex construction procedures, difficult multi-equipment coordination, and lack of real-time state perception in multi-face concurrent construction of underground cavern groups in water conservancy and hydropower projects, a multimodal perception-driven intelligent coordinated scheduling method for underground cavern groups is proposed. Targeting three key limitations of existing methods—insufficient perception robustness, disconnected perception-to-scheduling data pipeline, and weak adaptability to complex cavern topology—the innovations of this paper are as follows. First, video, audio, and IMU signals from a double-arm drilling jumbo are fused through adaptive quality gating and physical prior injection to construct a construction state recognition model (TMSR), achieving accurate and continuous identification of key operational states under harsh conditions. Second, based on the recognized process states, the Protodyakonov coefficient and bulking factor of surrounding rock are inversely estimated in a rolling manner using the drilling-time method and temporal smoothing algorithm, and a discrete-event simulation environment incorporating physical topological relationships is established to bridge the perception-scheduling data pipeline. Third, multiple dump trucks are abstracted as a heterogeneous multi-agent system, and a reconstructed dynamic active-fleet reward mechanism combined with a reinforcement learning algorithm realizes adaptive dispatching decisions across multiple working faces. Case study results demonstrate that the dynamically inverted geological parameters show high consistency with advance geological forecasts, with an average relative error of only 5.6%. Compared with on-site manual scheduling, the proposed strategy reduces total construction duration by 9.5% and decreases average dump truck queuing time by 6.3 minutes per trip. The proposed method provides significant engineering value for improving construction efficiency in underground cavern projects and advancing intelligent construction practice.