Multiple questions on velocity planning

[Kim-mins]

Hi team,

Thank you for the nice ADS!

I'm currently looking at the Autoware documentation, and I got some questions regarding velocity planning as below:

Q1. When the obstacles are ignored(no surrounding NPCS, ...), is it possible to infer a velocity at some point directly(=not running the driving scenario) with any of Autoware's component?
For this question, I focused on the data type Path and Trajectory, as they contain a velocity information.
So I mainly investigated behavior velocity planner and motion velocity planner, as they outputs those types of data, but when I checked the documentation of those components, inferring a velocity at some point seems not directly supported since they require information which can be obtained while running the driving scenario(e.g., sensor data, acceleration of the ego vehicle). Could you please tell me if my understanding is correct?

Q2. How much are the velocity data from behavior velocity planner and motion velocity planner accurate?
behavior velocity planner outputs data, typed Path, and it contains very raw velocity information, which is not accurate.
But when it comes to the data, typed Trajectory, from motion velocity planner, it seems the velocity value is accurate.
I wonder what those values are for, and how much they are similar to the velocity while running the driving scenario.

Thank you!

[maxime-clem]

Let me start by answering the 2nd question as it makes it easier to answer the 1st question.

Q2. How much are the velocity data from behavior velocity planner and motion velocity planner accurate?

The velocities in the Path message calculated by behavior planning correspond to upper bounds velocities. Therefore they cannot be considered accurate and the actual velocity profile of ego may greatly differ from these values.

The velocities in the Trajectory message calculated by motion planning correspond to target velocities that the controller should follow. With a perfect controller/vehicle, these velocities would be 100% accurate, but in practice there will always be some error.

Q1. When the obstacles are ignored(no surrounding NPCS, ...), is it possible to infer a velocity at some point directly(=not running the driving scenario) with any of Autoware's component?

The velocity_smoother (https://autowarefoundation.github.io/autoware.universe/main/planning/autoware_velocity_smoother/) is the module calculating the "accurate" velocity profile. However, as discussed in the reply to Q1, some error will be introduced by the controller and vehicle. It is thus difficult to precisely infer a velocity at some point. To observe the error issue, you can enable the View Text Velocity for the trajectory message in Rviz and see how it changes while driving.

simplescreenrecorder-2024-10-07_13.57.42.mp4

[maxime-clem]

If you want the Trajectory to match what would be calculated at runtime by Autoware then yes you need to run all planning nodes.

[Kim-mins]

Thank you for the clear answer! It was really helpful!

[Kim-mins]

Hi @maxime-clem!

I'm currently reading every your comment again carefully, and I just got more questions.. Could you please check it for me?

Q1. Regarding what you told me about some errors of control module, are they the direct cause of the discrepancy between the initial velocity plan and the real velocity while driving?
I mean, I want to check if those delay, resampling, precision and vehicle (and more) makes velocity while driving deviate from the initial velocity plan.

What I understood from your previous answers is, even though there's a initial velocity plan, those errors make the initial plan unachievable(=results in inaccurate accel/brake).
So the velocity at some point while driving differs from that of the initial plan, and the plan is updated accordingly, considering a current velocity and other states.

Q2. (This asks some more details of your answer) For the delay, you told me as the cause of the error, could you please tell me some details on how does delay make the error?
From your previous answer, what I understood is this: if the ego vehicle is accelerating from a starting point, and both the current and target velocities are 3.0 m/s, the velocity used for calculating the current acceleration could be 2.9 m/s, which is taken from a recent past state due to the delay. This results in a higher calculated acceleration, causing the vehicle to accelerate more, which in turn leads to the velocity to deviate from the initial plan. Is this correct?

Sorry for the long following questions. I'd really appreciate it if you could answer them for me.

Thank you!

[maxime-clem]

Sorry for the long following questions.

No problem at all ! These are interesting questions.

Q1

Your understanding seems correct. Here is a simplified example of the plan changes between iterations:

• t=0 : Autoware calculate a plan [p_t0, p_t1, p_t2, p_t3, ...].
  • p_ti is the state (position & velocity) of the ego vehicle at t=i.
  • p_t0 is the current state.
• t=1 : Autoware calculate a new plan [p_t1', p_t2', p_t3', ...].
  • p_t1' (current ego state) can have some error compared with the previously planned p_t1 and this new state is used for computing the new plan.

Q2

Your example of delay is correct.
The controllers have parameters to compensate for some constant delay but in practice some error always remains (e.g., because communication/calculation times are not constant).

[Kim-mins]

Thank you for the confirmation!
And I could get more details with your examples. Thanks a lot!