Planner does not react to dummies' predicted trajectory when on priority road

[LLusvarghi]

Checklist

• I've read the contribution guidelines.
• I've searched other issues and no duplicate issues were found.
• I'm convinced that this is not my fault but a bug.

Description

Hi,
I am using autoware.universe 2024.01 on Ubuntu 22.04 LTS.
I am currently running autoware on CARLA's Town01 map.

I noticed that the planner does not react (i.e., does not modify the currently planned trajectory) to the dummy vehicles' predicted trajectories when the dummy vehicles are on a road with lower priority. It reacts to a dummy vehicle only when it physically is on the currently planned trajectory and is therefore perceived as an "obstacle on the path".

Does anybody know why this is happening?

Expected behavior

The planner should plan a new trajectory to avoid the imminent collision with a dummy vehicle that has a predicted future trajectory that crosses the currently planned trajectory. This is what normally happens when the ego-vehicle is on a road with lower priority or when all roads have the same priority. See screenshot below.

Actual behavior

The planner does not take into account potential collisions that could occur with dummy vehicles that are driving on a lower priority road. It seems like that the planner is giving more importance to the road priorities and it assumes that the dummy vehicles will respect the driving rules and stop since they are on a lower priority road. The planner does not take into accoutn the dummy vehicles' predicted trajectories estimated by the perception module and reacts to the dummy vehicle only when it physically is on the currently planned trajectory. See screenshot below.

Steps to reproduce

1. Assign different priorities to CARLA's Town01 map roads
2. Launch the ego-vehicle on a higher priority road in correspondence of an intersection
3. Set the ego-vehicle's goal destination so that it will cross the intersection
4. Generate a dummy vehicle on a lower priority road such that it will cross the ego-vehicle's planned trajectory in correspondence of the intersection

Versions

No response

Possible causes

No response

Additional context

No response

[marvsee]

Hi @LLusvarghi

Thanks for raising this issue.

I'm not entirely sure what you mean by assigning different priorities to the roads. Are you referring to defining the right-of-way using a regulatory_element with right_of_way and yield attributes?
If so, I believe the "Actual Behavior" you described is the correct and intended functionality. If the planner were to yield to every predicted trajectory regardless of the right-of-way, the ego vehicle would almost always yield, even when it has priority.

In your example, this would mean the following: The ego vehicle is on a priority road and detects a predicted trajectory intersecting its path. It would then stop and yield, even though it has the right-of-way, and would only proceed after the other vehicle comes to a complete standstill (causing its predicted trajectory to disappear).
This is not how a human would drive, and it would be overly cautious. In my opinion, it's reasonable to assume that other road users will follow the rules, at least to a certain degree. Driving too defensively can also lead to unpredictable and unsafe behavior in traffic.

If you have a different opinion on that, I am happy to discuss it further with you :)

[LLusvarghi]

Hi @marvsee, thank you for your prompt reply!

We are not actually using the right_of_way and yield attributes to define the road priorities. We are just setting the turn_direction tag and I understand that this is sufficient to let autoware understand road priorities. I am attaching Town01's lanelet file (in txt format) to this message. The IDs of the lanelets involved in the issue mentioned above are 29086 and 28984.

I understand the point of your reply, and I partially agree with you. I understand that "if the planner were to yield to every predicted trajectory regardless of the right-of-way, the ego vehicle would almost always yield, even when it has priority" and I agree that this would not be a very human-like driving behaviour.
Yet, I think that there might be corner cases (such as the one reported in my issue) where an imprudent vehicle (or an emergency vehicle, for example) on a non-priority road might not respect road priorities and that the ego-vehicle should be able to detect and react to such abnormal behavior. For example, I think that the ego-vehicle should be able to understand when a vehicle approaching from a non-priority road will not have sufficient time (considering the braking capabilities of a standard vehicle, for example) to decelerate and stop before the intersection. In my opinion, it should not completely disregard its predicted trajectory and react only when the vehicle is exactly on its planned trajectory.

In summary, I believe that a more "normal" behavior could lie between the two extremes: reacting to every predicted trajectory and not reacting to any predicted trajectory. When the ego has the priority, predicted trajectories should be taken into account when there is no physically feasible way for the approaching vehicle to decelerate and give priority to the ego.

But this is just my opinion. Looking forward to hearing your feedback!

Town01's lanelet file:
lanelet2_map.txt

[marvsee]

Hi @LLusvarghi

Thanks for the clarification. I understand your point now, and I agree. The ideal behavior is to react to a vehicle that is clearly not going to stop, even if we have the right-of-way.

You made a great point here:

For example, I think that the ego-vehicle should be able to understand when a vehicle approaching from a non-priority road will not have sufficient time (considering the braking capabilities of a standard vehicle, for example) to decelerate and stop before the intersection.

I think to achieve this, the system would require a more advanced prediction module that can distinguish between a vehicle that is decelerating to stop and one that will run through the intersection. Or it would need a collision check that considers physical braking capabilities, not just the positions and standard predictions.

But as far as I am aware, these specific capabilities are not currently implemented in Autoware. But I agree, that this would make sense for future developments.

[LLusvarghi]

Hi @marvsee,

Thank you very much for the support and the insightful discussion.
Do you know if, by any chance, this kind of feature is supported (actually, learned) by the new AI planner recently released by autoware?

[idorobotics]

Hi @marvsee,

Thank you very much for the support and the insightful discussion. Do you know if, by any chance, this kind of feature is supported (actually, learned) by the new AI planner recently released by autoware?

@m-zain-khawaja is this feature supported or expected soon?