ReVP: Incorporating Rotation Equivariance to Boost Spatio-Temporal Swarm Motion Prediction

Swarm motion prediction is crucial for understanding and controlling complex multi-agent systems. In the field of spatio-temporal predictive learning, convolutional models offer a favorable trade-off between accuracy and computational cost, regarded as a viable alternative to recurrent models. This paper proposes ReVP, a novel Rotation Equivariant Video Prediction model that leverages rotation equivariant convolutions to capture spatio-temporal dependencies under geometric transformations. By incorporating rotation equivariance, models benefit from enhanced data efficiency and training stability, along with parameter reduction due to the parameter sharing mechanism. Experiments on the Swarm Motion dataset show that ReVP outperforms baselines in prediction accuracy while requiring fewer parameters. Ablation studies validate the advantages of rotation equivariant models on different groups, including scenarios with and without data augmentation.