About training checkpoint

[Trunk2004]

Hi @Kait0 , I just want to ask that it usually take how long for the agent be able to steer ?
Mine take only 3 hours for a decent driving except for steering. Can you recommend which town should I use that have lot of steering scenarios?

Also, did you approximate the purple lane for the agent to make turn or the agent solve it all by it self and the purple lane just for termination if the agent is run off route?

Tks you and sorry for asking too much.

[Trunk2004]

Hi, does the use_extra_control_input have error when training ? When evaluate, I can't get the visualize due to error:

Actor could not be found in the registry. Actor Id: 2377

The traceback said it was from the use_extrac_control_input block in the env_agent

[Kait0]

use_extra_control_input is an experimental feature that we did not end up using.

[Kait0]

The purple lane are the lane markings of the street. They come from the HD map

[Trunk2004]

Tks you.
It weird to ask, but in your experiences with CARLA, can you recommend me which town map that can help the agent to steer/turn properly?

[Trunk2004]

Also @Kait0 , does the simple_reward traffic light detector match with the leaderboard red light test ? I see it registering to the terminal 3-4 red light running when evaluate my model but the leaderboard test only recognize 1-2 of them

[Trunk2004]

And about use_ttc, if I set it True then do I need to modify the value_head for it to receiver the remaining_ttc_penalty_ticks? Or the remaining_ttc_penalty_ticks is just an additional value for value network and it won't affect the reward computation?
Tks.

[Kait0]

no the reward traffic light detector in simple reward also counts red light infractions when the agent is invading the opposing lane while entereing the intersection while the metric only counts RL infractions in the agents own lane (a bug that the RL agent is exploiting if not fixed in the reward).

[Trunk2004]

Tks @Kait0 , that explain why the red light infractions count is different, so when training the reward still do it work and terminate the eps if it run red light to prevent exploiting but when evaluate, the agent can exploiting that bug and make some risky behaviors.

And about use_ttc, if I set it True then do I need to modify the value_head for it to receiver the remaining_ttc_penalty_ticks?

The ttc is work as the extra penalty to prevent the agent collision combine with termination penalty, right? And about the timeout, does it was the number of ticks each eps or it maybe different because of system time and game time in CARLA? As there are many eps which have length above the 1200.0 eval time.

the reward traffic light detector

So the detector is able to cover both lanes, I wonder that if I accidentally adding another red light check but it come from the carla.TrafficLightState() combine with the self.vehicle.is_at_traffic_light()
Then the agent is actually exploiting the bug ?