Create reward function for RL

[bcaine]

Potentially:

if throw:
    return - distance - max(distance)
else:
    return distance

Do we need a differentiable one instead?