Researchers introduce FlowMoP, a conditional flow matching framework that predicts diverse future motions for multiple people in 3‑D. The system jointly models global trajectories and root‑relative poses within a single generative model, enabling richer motion representations compared to prior deterministic approaches.

FlowMoP compresses future motion into a small set of B‑spline control points, reducing computation while preserving smoothness. A dual‑stream transformer attends separately to observed history and noisy future tokens, and conditions each agent’s predictions on the motion of its neighbors, allowing interaction‑aware forecasting.

On the CMU‑Mocap (UMPM) and 3DPW datasets, FlowMoP’s best‑of‑six predictions achieve lower joint position error, average pose error, and final displacement error than deterministic baselines.

On WorldPose soccer data, FlowMoP establishes a new baseline, outperforming prior state‑of‑the‑art methods T2P and EMPMP in all evaluated horizons.

These results demonstrate that integrating flow matching, spline reparameterization, and interaction conditioning produces more accurate and diverse predictions for multi‑person 3‑D motion.