Finished Ackerman 1 pose
Index
New states
Before, we treated the distance to the target as an absolute value, which made sense because with a holonomic architecture the vehicle can reduce both distances simultaneously. In the new scenario, this is not the case, and for that reason, we take into account that it is not the same to be at an absolute 1m being vertically as it is horizontally.
1
2
3
4
5
6
7
return (
discretized_x,
discretized_y,
self.discretize(heading_error, step=0.1, max_value=np.pi),
discretized_speed,
discretized_min_distance
)
In addition, the way in which we treat the minimum distance has been modified. Before, it was assigned a sign depending on whether you were in front of or behind the vehicle, which makes little sense, as it can also be laterally. Now the value is absolute and the officer will have to learn whether to brake, turn, accelerate, etc., taking into account his or her speed.
1
discretized_min_distance = self.discretize(min_distance, 0.1, 1.5)
Heading problem
For several weeks the main problem we had (without being identified), is that the officer always tried to park in front of us. We tried shortening the front turning range, and the result was that he didn’t learn anything at all.
To address this, the reward system has been modified by making the local minimums much lighter and rewarding the vehicle for approaching the target backwards.