AI learns to cross the road with Unity ML-Agents (imitation learning or reinforcement learning).
Goal: Cross to the other side of the road without colliding with a car.
Agents: The environment contains 30 agents of the same kind, all using the same Behavior Parameters.
Agent Reward Function:
+5 for crossing the road.
-5 if the agent collided with a car.
+0.05 if the current distance is less than the last.
-0.005 if the current distance is greater or equal to the last one.
Behavior Parameters:
Vector observation space: 8 variables corresponding to rotation of the agent's cube and the position and distance to the other side of the road.
Size of 40, corresponding to 20 ray casts each detecting 2 possible objects.
Actions: 2 discrete action branches:
Forward Motion (3 possible actions: Forward, Backwards, No Action)
Rotation (3 possible actions: Rotate Left, Rotate Right, No Action)