This paper presents the experimental realization of a high-fidelity CZ gate based on a double-transmon coupler (DTC). The DTC scheme enables both suppressed residual ZZ interaction and a fast high-fidelity two-qubit gate, particularly for highly detuned qubits. By leveraging advanced fabrication techniques and a model-free pulse-optimization process based on reinforcement learning, the authors achieve a CZ gate fidelity of 99.92% and single-qubit gate fidelities of over 99.98%. The DTC scheme demonstrates its potential as a competitive building block for superconducting quantum processors. The DTC consists of four transmons: two data qubits and two coupler transmons. The ZZ interaction between the data qubits can be controlled by the magnetic flux through the DTC's loop and suppressed completely even for qubits outside the straddling regime. The absence of direct capacitive coupling between the qubits reduces spectator error and allows flexible qubit-qubit distances, minimizing both quantum and classical crosstalk. The authors overcome challenges in achieving a high-fidelity CZ gate experimentally by designing the chip for high coherence and fabrication feasibility. They achieve a high on-off ratio of the ZZ interaction between the two qubits, crucial for high gate fidelities. At the idle bias point, a minimum residual ZZ interaction of 2π×6.3 kHz persists without compromising single-qubit gate fidelities. They calibrate the distortion of the Z-pulse applied on the DTC using the readout resonator of a qubit and further optimize the pulse using a model-free reinforcement learning algorithm, achieving a fast and optimized pulse for a high-fidelity CZ gate with a fidelity of 99.92±0.01%. The CZ-gate fidelity, demonstrated through Clifford interleaved randomized benchmarking, remains stable within a 12-hour timeframe. The primary contributors to the CZ gate error are leakage and incoherent errors. The DTC scheme offers advantages such as less frequency-collision probability, flexible spatial arrangement of qubits, and simplified control degrees of freedom. These attributes make it highly promising for implementing NISQ applications and quantum error correction in the near future.This paper presents the experimental realization of a high-fidelity CZ gate based on a double-transmon coupler (DTC). The DTC scheme enables both suppressed residual ZZ interaction and a fast high-fidelity two-qubit gate, particularly for highly detuned qubits. By leveraging advanced fabrication techniques and a model-free pulse-optimization process based on reinforcement learning, the authors achieve a CZ gate fidelity of 99.92% and single-qubit gate fidelities of over 99.98%. The DTC scheme demonstrates its potential as a competitive building block for superconducting quantum processors. The DTC consists of four transmons: two data qubits and two coupler transmons. The ZZ interaction between the data qubits can be controlled by the magnetic flux through the DTC's loop and suppressed completely even for qubits outside the straddling regime. The absence of direct capacitive coupling between the qubits reduces spectator error and allows flexible qubit-qubit distances, minimizing both quantum and classical crosstalk. The authors overcome challenges in achieving a high-fidelity CZ gate experimentally by designing the chip for high coherence and fabrication feasibility. They achieve a high on-off ratio of the ZZ interaction between the two qubits, crucial for high gate fidelities. At the idle bias point, a minimum residual ZZ interaction of 2π×6.3 kHz persists without compromising single-qubit gate fidelities. They calibrate the distortion of the Z-pulse applied on the DTC using the readout resonator of a qubit and further optimize the pulse using a model-free reinforcement learning algorithm, achieving a fast and optimized pulse for a high-fidelity CZ gate with a fidelity of 99.92±0.01%. The CZ-gate fidelity, demonstrated through Clifford interleaved randomized benchmarking, remains stable within a 12-hour timeframe. The primary contributors to the CZ gate error are leakage and incoherent errors. The DTC scheme offers advantages such as less frequency-collision probability, flexible spatial arrangement of qubits, and simplified control degrees of freedom. These attributes make it highly promising for implementing NISQ applications and quantum error correction in the near future.