Composable Action-Conditioned Predictors: Flexible Off-Policy Learning for Robot Navigation | Read Paper on Bytez