A Hyper parameters and finer experimental details

The hyper-parameters used for our algorithm are shown in Table 1. The'Point' robot has steering and throttle as action space while'Car' robot has differential control. We use Performance Ratio (PR) threshold of 66%. Minimum 4 GB GPU space is required for running both the model based approaches. We compare how model learning validation loss varies in Safe RL setting as opposed to unconstrained RL one.