A new method called ASIR uses an inverse dynamics model to keep offline reinforcement learning agents within familiar state ...