We focus on the task of future frame prediction in video governed by\nunderlying physical dynamics. We work with models which are object-centric,\ni.e., explicitly work with object representations, and propagate a loss in the\nlatent space. Specifically, our research builds on recent work by Kipf et al.\n\\cite{kipf&al20}, which predicts the next state via contrastive learning of\nobject interactions in a latent space using a Graph Neural Network. We argue\nthat injecting explicit inductive bias in the model, in form of general\nphysical laws, can help not only make the model more interpretable, but also\nimprove the overall prediction of model. As a natural by-product, our model can\nlearn feature maps which closely resemble actual object positions in the image,\nwithout having any explicit supervision about the object positions at the\ntraining time. In comparison with earlier works \\cite{jaques&al20}, which\nassume a complete knowledge of the dynamics governing the motion in the form of\na physics engine, we rely only on the knowledge of general physical laws, such\nas, world consists of objects, which have position and velocity. We propose an\nadditional decoder based loss in the pixel space, imposed in a curriculum\nmanner, to further refine the latent space predictions. Experiments in multiple\ndifferent settings demonstrate that while Kipf et al. model is effective at\ncapturing object interactions, our model can be significantly more effective at\nlocalising objects, resulting in improved performance in 3 out of 4 domains\nthat we experiment with. Additionally, our model can learn highly intrepretable\nfeature maps, resembling actual object positions.\n