Hi,
We noticed that the heading of our vehicle takes values between 0 and 360, while the heading of the social vehicles takes values between -360 and 0.
Additionally, we noticed that even though our vehicle and the social vehicles are heading in the same direction, their heading values differ by 90 degrees.
Could you please fix this issue?
Thank you
Hi,
Thanks for your question. That's strange - I just tested this now and the heading of ego and neighborhood is computed using the same coordinate frame and so I don't see this difference. Are you able to share the exact calculation you are using to determine this discrepancy? Is it possible you are mixing relative coordinates with absolute?
I tested by trying variations on:
if len(env_obs.neighborhood_vehicle_states) > 0:
print(env_obs.neighborhood_vehicle_states[0].heading)
print(env_obs.ego_vehicle_states.heading)
With a single social vehicle in the environment.
Posted by: HuaweiUK @ Dec. 4, 2019, 10:57 a.m.Hi,
This is very interesting. We are using exactly the same code you posted but we notice the discrepancy we mentioned. We are currently using sumo 1.3.1 (since the older one was crashing).
Are you maybe using the 1.2 version? It is important that the evaluations and the training are done with the same version.
Also, was the starter kit (hiway python package) ever updated?
Thanks.
Hi, we were able to reproduce this behavior and unfortunately it's a difficult one to fix since other contestants likely have models that depend on the existing behavior.
For now we can only recommend using this transformation function to map the neighborhood vehicle headings into the same space as the ego heading.
https://gist.github.com/davidrusu/d144a2646c9597a0d412c7e795ada548#file-nv_heading_to_ego_heading-py
The evaluation system is indeed running the same code you are and so your model will see the same heading angles in evaluation as you do in training.
Posted by: HuaweiUK @ Dec. 6, 2019, 5:09 a.m.