Actions

[vcoyette]

The environment used in the launch_env function is of type Simulator, which means that the actions passed to the env correspond to the wheels velocity (left and right).
The ActorCNN network of the reinforcement training uses a sigmoid on the first action, which acts on the left wheel, thus the resulting left wheel velocity is between 0 and 1 but the right wheel velocity is between -1 and 1.
The sigmoid on the first action makes sens if the actions were, as described in the README, the velocity and the steering angle, but this is not the case. Testing the enjoy_reinforcement with a policy giving constant action of [1, 0] can convince yourself of this behaviour : if the README were up to date, the duckie should go forward. If you try this policy, it turns continuously.

Thus the README may need and update as well as the actorCNN.