In view of the key problems of the current virtual character dynamic performance in complex interactive scenes, such as insufficient body coordination and distortion of motion details, this study proposes a dynamic performance optimization framework based on multimodal motion capture data fusion and physical constraint optimization. First, this paper constructs a multidimensional dataset containing body motion, joint torque and environmental interaction force. Then, this paper designs a feature extraction network based on the temporal attention mechanism to achieve the decoupled representation of action semantics and physical parameters. Finally, a hierarchical physical constraint optimization strategy is proposed to ensure the physical rationality of virtual characters in complex interactive scenes such as collision and weight bearing. The experimental results show that compared with the Vicon optical capture system, the optimized framework reduces the joint angle error by about 42.7%; in terms of interactive experience in virtual reality, the optimized framework scores 7.5; in terms of motion redirection accuracy, the error range of the optimized framework is 0.68 m to 0.85 m. These experimental results verify the effectiveness of the framework based on multimodal data fusion and physical constraint optimization in improving the accuracy and naturalness of virtual character dynamic performance, and provide technical support for high-quality character generation in virtual reality.