Numerical analysis

From our experiments we concluded that the bird's-eye view offers better performance than the visual input (both for NGSIM and openDD). More precisely, we evaluated both the polices on the same set of test scenarios. Importantly, the test scenarios were not used during the training, so our evaluation tests generalization of RL trained polices.

Interestingly, for NGSIM the bird's-eye is 'uniformly' better.

Multiple seeds

In order to measure stability of training, we run 3 copies of each of the main experiments. Below are the result of evaluation of the respective models.