dynamics-aware

DynaFLIP is a dynamics-aware multimodal pre-training framework that integrates motion understanding into visual perception for robot manipulation. It uses image-language-3D flow triplets and geometric regularization to improve representation learning, achieving significant gains in out-of-distribution scenarios.