dynamic_mpc.py causes crash when using it with a different map

[gurtajbir]

Hello.

I wanted to try out the dynamic_mpc.py with some other map than the levine map. I tried using it with the Austin map by using the corresponding files from here . I had a similar experience with the Spielberg map as well. The simulation starts out OK but soon the vehicle veers off the raceline and crashes into the boundary at which point the simulation ends. The attached picture is a screen capture just before the collision happens.

Has anyone encountered this issue and been able to successfully mitigate? If yes, could you provide some suggestions?

Thanks!

[gurtajbir]

@luigiberducci @hzheng40 @AhmadAmine998 Any suggestions?

[AhmadAmine998]

Which branch are you using? I'll try to replicate this

[gurtajbir]

@AhmadAmine998 I used the main branch.

[gurtajbir]

@AhmadAmine998 Did you get a chance to take a look at this?

[AhmadAmine998]

@AhmadAmine998 Did you get a chance to take a look at this?

Apologies for the late reply, I just had the chance to test it. Did you make sure to configure the dlk for this set of waypoints? It seems from your photo that your reference (purple x) is much more spread out than your predicted which might indicate a mismatch between your waypoints' actual dlk and the config dlk. Try using the following:
planner.config.dlk = np.linalg.norm(np.array([waypoints[0, 1], waypoints[0, 2]]) - np.array([waypoints[1, 1], waypoints[1, 2]]))
planner.config.dl = np.linalg.norm(np.array([waypoints[0, 1], waypoints[0, 2]]) - np.array([waypoints[1, 1], waypoints[1, 2]]))

I looked into the controller performance in detail and it seems the current implementation is not very stable. I have taken note of it and added it as a new issue.

Do note that the main branch is quite outdated and we are working towards a much cleaner implementation that addresses many bugs and pain points in this old implementation. If you just need the code for a reference MPC implementation, I would highly suggest you check out v1.0.0. Note that it will require you to use v1.0.0 for the gym environment as well. I can not confirm if this is an issue on v1.0.0, but the code should be better than this.