Motion prediction is crucial for autonomous driving systems to understand complex driving scenarios and make informed decisions. However, this task is extremely challenging due to the diversity of traffic participant behaviors and the complexity of the environment. Existing edge prediction methods assume that the agents' future distributions are independent, thus ignoring the interactions between future trajectories. We propose a joint prediction strategy that incorporates the theory of traffic participant risk quantification to achieve scenario-level optimization of agent trajectory distributions. Meanwhile, in order to enhance the scenario compatibility of anchor-free decoding methods and avoid the over-reliance of anchor-based methods on the quality of manually selected targets, we employ learnable queries as adaptive reference information in the trajectory decoder. The proposed multi-agent joint prediction framework based on risk field theory (RMP) enables the predicted trajectories to be more compatible with actual traffic behavior and possesses good robustness in complex scenarios. We validate the performance of the model on the publicly available Argoverse1 dataset. Experiments show that RMP performs competitively on predictive metrics, achieving the best ADE1 (1.579) and DAC performance (99.1%) compared to other state-of-the-art models. This study provides technical support for accurately predicting the motion of autonomous driving systems.