Injecting knowledge in data-driven vehicle trajectory predictors

Vehicle trajectory prediction tasks have been commonly tackled from two distinct perspectives: either with knowledge-driven methods or more recently data-driven ones. On the one hand, we can explicitly implement domain-knowledge physical priors such as anticipating that vehicles will follow middle of roads. While this perspective leads to feasible outputs, it has limited performance due difficulty hand-craft complex interactions in urban environments. other recent works use approaches which learn data leading superior performance. However, generalization, \textit{i.e.}, having accurate predictions on unseen data, is an issue unrealistic outputs. In paper, propose a "Realistic Residual Block" (RRB), effectively connects these perspectives. Our RRB takes any off-the-shelf model and finds required residuals add knowledge-aware trajectory. proposed method outputs realistic by confining residual range taking into account its uncertainty. We also constrain our output Model Predictive Control (MPC) satisfy kinematic constraints. Using publicly available dataset, show outperforms previous terms accuracy generalization new scenes. release code split here: https://github.com/vita-epfl/RRB.