Investigate incorporating more lidar into the state space

[retinfai]

Currently, we are reducing the 640 points of lidar down to 10 by averaging. A great deal of information is lost which could help the agent drive better.

The reason for the current approach is learning time. 1M steps is currently what is standard. A previous experiment using CNNs resulted in the car driving only ~30 steps before crashing after 500k steps. Thus, a great deal more training in the order of 10M steps would be necessary to train the agent.

Investigate this more thoroughly. One idea, is to use CNN layers to reduce the lidar signal before passing it through fully connected layers to determine the final actions of the agent.

[retinfai]

additionally, maybe training for 10M steps is also an approach