AI
New AI Model Predicts Human Movement 100× Faster Using One-Step Flow Matching
A novel machine learning technique called MoFlow combines flow matching with knowledge distillation to forecast pedestrian and athlete trajectories with unprecedented speed and accuracy.
Forecasting where people will move next—whether on a crowded street or a basketball court—remains a challenging task for machine learning systems. Researchers have now introduced MoFlow, a new approach that addresses this problem by generating multiple plausible future paths while running dramatically faster than previous methods.
The Challenge of Multi-Modal Prediction#
Human movement is inherently unpredictable because the same starting situation can lead to many different outcomes. A pedestrian approaching an intersection might turn left, go straight, or wait. The goal of trajectory forecasting is to generate K different future paths that capture this natural variation while remaining physically and socially believable. Existing flow-matching models can do this well but require many sampling steps—making them slow for real-world applications.
MoFlow's Novel Approach#
The researchers developed two key innovations. First, they designed a specialized flow matching loss function that ensures at least one of the K predicted trajectories is accurate while pushing all of them to be diverse and realistic. This balances quality with variety in predictions.
Second, they applied implicit maximum likelihood estimation (IMLE) as a distillation strategy. Rather than training a lightweight "student" model directly, knowledge is transferred from the full teacher model using only samples it generates. This approach proved far more effective than conventional distillation methods.
Impressive Performance Gains#
Testing on real-world datasets—including NBA SportVU games, ETH-UCY pedestrian sequences, and Stanford Drone Dataset footage—both the full MoFlow model and its distilled version achieved state-of-the-art results. The one-step distilled student runs 100 times faster than the teacher model during inference while maintaining competitive accuracy.
The team has released code, trained models, and data to support reproducibility and further research.
