Too optimistic behaviour of the "CC human driver" for the ALKS deceleration scenario?

[mdeppe]

I'm currently inspecting the code for the deceleration behaviour of the "CC human driver" for the ALKS deceleration scenario. I took the Reg 157 and the JAMA paper (https://www.grcc.vip/article-7247.html) as a reference and tried to match the diagrams for the "skilled human performance model" with the calculations and parameters in the Python code (models.py, vehicle.py, movement.py, ...).
So I took this as a reference:

I realized to things:

1. in the code I did not see the risk perception time of 0.4 seconds. In the code the CC human driver immediately starts to brake for 0.75 seconds with 0.4m/s^2 deceleration and then start to ramp deceleration to 0.774G. See the green line in the picture below.
   Expectation (see red line in the picture below) is to have acceleration 0.0 for 0.4 seconds after the leading vehicle reaches 5m/s^2 deceleration. Then 0.75seconds with acceleration 0 and a transition to coast state (the vehicle rolls with negative engine torque, rolling resistance and air drag) with a small deceleration. Then after 0.75 seconds ramping down to 0.774G.
2. The code does not detect if the leading vehicle reaches the 5m/s^2 deceleration. If one f.ex. provides a deceleration of 3m/s^2 as command line argument to the "run_one_case" the human driver starts to decelerate with 0.4m/s^2 immediately despite the leading vehicle never is at 5m/s^2.

When comparing the red and green lines the green one is a more optimistic performance model for the human driver. It decelerates early and does not consider acceleration 0 where the driver still has the foot on the accelerator pedal. (Acceleration 0 means the engine sill produces positive torque and compensates the driving resistances). Could you please comment.

[mdeppe]

@kostasmatt @ricDona93 @gabbaan : Is this still an active project? How can I reach the model developers?

[RicDona-JRC]

Hello @mdeppe, I contributed to the development of the project as @ricDona93 but now I have a new account due to a new e-mail address.

I shall start by saying that the EC-JRC did not develop the CC model but we provided a simulation implementation following extensive conversations with the developers to make sure that the code implementation was correct.

On the other points you rightfully raised:

• the rolling/aerodynamics resistance and negative idle torque are not modeled in any of the models as they are very much vehicle-dependent and given the high abstraction level of the models themselves. Anyway, via using this approach the models are "less" performing hence they don't imply unreasonable performance;
• the 0.4s "risk perception" is implemented in line 92, file: JRC-FSM/utility/models.py;
• the deceleration magnitude reaction produced by the ego-vehicle is not leader-dependent except for the FSM. All the other models will use either full deceleration or none as per the specifications.

I hope this clarifies.

Cheers
Riccardo

[mdeppe]

Hello @RicDona-JRC ,
thanks for commenting this. I think we are not yet fully aligned. So my first question ist: Are you interested in improving the CC Human Driver model code of the repository? I volunteer to support this.

Here the technical stuff as comments to your response:

• the rolling/aerodynamics resistance and negative idle torque are not modeled in any of the models as they are very much vehicle-dependent and given the high abstraction level of the models themselves. Anyway, via using this approach the models are "less" performing hence they don't imply unreasonable performance;
  =MD=> At least there is something implemented for the cc human driver as a constant deceleration of 0.4m/ss. This seems to be a kind of guess (according to the comment):
  

• the 0.4s "risk perception" is implemented in line 92, file: JRC-FSM/utility/models.py;
  =MD=> To my understanding the code around line 92 implements the braking delay of 0.75 seconds:
  
  

• the deceleration magnitude reaction produced by the ego-vehicle is not leader-dependent except for the FSM. All the other models will use either full deceleration or none as per the specifications.
  =MD=> Here my point is that I have the feeling that the "risk perception" time of 0.4sec for the cc human driver is not implemented (see the discussion point above). If one would implement the "risk perception" time it is necessary to have a trigger to define where risk perception starts. In the ALKS spec this is when the leading vehicle reaches 5m/ss deceleration.

[RicDona-JRC]

Hi @mdeppe, ok now I think I got your point better.

I was initially confused by the 0.4s "risk perception" time and 0.4 m/s2 corresponding deceleration and I was not completely sure to which quantity you referred. The 0.4 m/s2 braking due to throttle pedal release is implemented in the mentioned line of code following the 0.75s delay that you rightfully pointed out.

You're of course welcome to download the project and provide your own implementation of the code based on the interpretation of the JAMA document.

Please notice that the CC model in the ALKS is defined in the regulation itself in Annex 3:
https://unece.org/sites/default/files/2023-03/R157am4e%20%281%29.pdf

I am not entirely sure whether the model in the position paper you shared is exactly the one implemented in the legislation as I was not the one who crafted the CC model.

As I mentioned, we developed, implemented, and validated the FSM model only(https://www.sciencedirect.com/science/article/pii/S0001457522001798#f0005) whereas we only provided a numerical implementation for the CC which we agreed upon together with the corresponding developers.

The results that this repo produces are totally comparable with the charts in UN-R157 so in case any discrepancy exists between the way we understood the CC model and what is envisaged in the position paper it should be a minor one.

On the 5.0 m/s2 leader deceleration effect, this parameter is only relevant for the 'Deceleration' scenarios where all collisions should be avoided as foreseen by the CC model in the regulation and predicted by the numerical implementation in this repo.

I hope this helps.
Riccardo

[mdeppe]

Hi @RicDona-JRC,
thank you, I think we are coming closer to the point. In fact I'm referring to the exact CC model as specified in the ALKS regulation in Annex 3 in all my comments. And you're right, I'm referring to deceleration scenarios only with the focus on the CC model only.
This is the focus of performance model 1: "ALKS p. 40, 3.3.1" : "In the first performance model, the avoidance capability of the driver model is assumed to be only by braking."

When running the cc model from the repository with: "one_case", "car_following", "CC_human_driver" it produces this:

So my point is that the acceleration time plot in the "Acceleration" figure does not match with the characteristics of the accelerator and brake pedal timing as shown in ALKS p. 41 Figure 1: "Competent and careful human performance model".
The 0.4s "risk perception" time is missing and the acceleration is not smooth which will produce a lot of jerk.
I definitely see room for improvement.
Best regards,
Markus

[kostasmatt]

Hi Markus,

Thanks a lot for reaching out.
Indeed, when trying to simulate according to the CC human driver model, the breaking scenario has been challenging. As you point out there is: "The risk perception time [a] is 0.4 seconds. The risk perception time [a] begins when the leading vehicle exceeds a deceleration threshold 5m/s2."
However, in the regulation document, page 56, we see results using leader decelerations that are smaller than 5 m/s2. In such cases, if we would add the 5 m/s2 deceleration threshold, and started counting 0.4 seconds after it, we would never have any reaction.
Nevertheless, this is one of the most straight forward cases, regarding the regulation, as the ALKS should avoid the accident with any leader deceleration. If it could not, it should maintain a longer steady state distance. So the use of CC human driver for such cases may be redundant. Thus, we opted to focus more on the other two cases, that are more relevant to setting up performance requirements.

You also mention that "the acceleration is not smooth which will produce a lot of jerk". You mean when the vehicle comes to a stop and the acceleration goes to 0? If that is the case, I am not sure that it would affect the accident avoidance. The deceleration is bound the the minimum jerks agreed upon.

Best Regards,
Kostas

[mdeppe]

Dear @kostasmatt ,
you wrote: Indeed, when trying to simulate according to the CC human driver model, the breaking scenario has been challenging. As you point out there is: "The risk perception time [a] is 0.4 seconds. The risk perception time [a] begins when the leading vehicle exceeds a deceleration threshold 5m/s2."
However, in the regulation document, page 56, we see results using leader decelerations that are smaller than 5 m/s2. In such cases, if we would add the 5 m/s2 deceleration threshold, and started counting 0.4 seconds after it, we would never have any reaction.
=MD=> Good point, seems to be a weakness in the ALKS regulation: What is the risk perception starting point for the braking scenarios with leading vehicle decelerations smaller than 5m/ss?

you wrote: Nevertheless, this is one of the most straight forward cases, regarding the regulation, as the ALKS should avoid the accident with any leader deceleration. If it could not, it should maintain a longer steady state distance. So the use of CC human driver for such cases may be redundant. Thus, we opted to focus more on the other two cases, that are more relevant to setting up performance requirements.
=MD=> So you decided to leave out the risk perception time for all cases. This means even for the cases where the leading vehicle deceleration exceeds 5m/ss!
This situation is in the plots below. It is "one_case", "car_following", "CC_human_driver" with 0.6*9.81m/ss (5,89m/ss) deceleration for the leader.
I added a plot to your code for the leading vehicle deceleration.
In my eyes this is a highly relevant scenario BUT the risk perception time is not modelled meaning that the specification for the CC human driver is not met. Result is a too optimistic driver behaviour.
All in all this at least is worth to be mentioned in some kind of documentation.

you wrote: You also mention that "the acceleration is not smooth which will produce a lot of jerk". You mean when the vehicle comes to a stop and the acceleration goes to 0? If that is the case, I am not sure that it would affect the accident avoidance. The deceleration is bound the the minimum jerks agreed upon.
=MD=> True, I added the plotting of jerk to your code. In fact coming to standstill produces a peak. I agree that this is irrelevant.

=MD=> I was more referring to the first 750ms. Here we see zero jerk and constant -0.4m/ss acceleration in the plots but the ALKS regulation describes foot transfer from accelerator to brake pedal. Then it takes another 0.6seconds to reach 0.774G which looks correct in your model (note: the JAMA figure below wrongly shows 0.774sec instead of 0.774G).
I would expect some slightly different (non constant) jerk/accel curves in the first 750ms.
Figure 8 in https://www.grcc.vip/article-7247.html