Attempt to use only corrective steering for steering.

[vguerra]

Let's see how the test rosbag behaves when using only corrective steering.

108

[asimonov]

Victor,

Ok, i understand now. No, i did not run it yet.

Just thinking now if twiddle will actually work in this case... the commands we send when playing rosbag are not actually affecting the pose we receive, so will our twiddle have the right error to tune?

---

Kind Regards, Alexey Simonov

On 13 Nov 2017, at 08:58, Victor Guerra notifications@github.com wrote:

@vguerra commented on this pull request.

In ros/src/twist_controller/twist_controller.py:

@@ -52,7 +52,7 @@ def init(self,

current kp, ki and kd are values found from previous twiddle runs.

self.steer_pid = PIDWithTwiddle("steering PID", kp=0.607900, ki=0.000172, kd=1.640951, mn=-max_steer_angle, mx=max_steer_angle,

• optimize_params=False, iterations=10, tolerance=0.05)
• optimize_params=True, iterations=10, tolerance=0.05) we have two options here:

Have a run with optimize_params=True on test rosbag and use those parameters for the resubmition ( only if we see an improvement on the test rosbag, .i.e. the car follows waypoints ) and optimize_params=False. We simply go ahead and resubmit with optimize_params=True and let the algo learn the optimal parameters on the field. Did you have the chance to test this locally using the test rosbag?

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.

[vguerra]

yes.. i hope so.

My assumption is that steering command should be more aggressive.

If car still does not follow properly the waypoints with optimize_params=True, then the issue lies somewhere else.

[asimonov]

but I do not think the car would respond in rosbag play mode. our ROS code will be sending steering commands, but they do not affect pose that it receives back. it will receive pre-recorded pose, I would imagine. so I think the twiddle will be trying to tweak parameters without any response in CTE and the resulting parameters will be extreme. or am I missing something?

On 13 Nov 2017, at 12:27, Victor Guerra notifications@github.com wrote:

yes.. i hope so.

My assumption is that steering command should be more aggressive.

If car still does not follow properly the waypoints with optimize_params=True, then the issue lies somewhere else.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/Kairos-Automotive/carla-brain/pull/109#issuecomment-343904163, or mute the thread https://github.com/notifications/unsubscribe-auth/AAOr9m9X-8k15SICjEXVA9sAU26BOLWtks5s2DWigaJpZM4Qak8v.

[yhoazk]

I believe the same as @asimonov, the twiddle won't work with the rosbag as the feedback is fixed. I'll work tonight in creating the waypoints which are closer to the Carla test track, tight curves and and just one loop. Then we can use this waypoints in the simulator, and there the twiddle will be able to adjust the parameters accordingly to a tighter turn.

[vguerra]

ah ok! .. yes.. i get it now .. silly me :p

you are both right ... it is kind of useless to test this against the rosbag we got as feedback.

so yes.. we could still check in the simulator how the car behaves with only PID/Twist controller.

I remember I introduced the twiddle algorithm after we started to use corrective and predictive steering. I never tested steering with PID alone, probable performs better :).

[asimonov]

Vector, we can get rid of predictive steering anyway. can you just disable True for tuning parameter?

Gustavo, that's an interesting idea. So the car will just run in a small circle on the grass next to the actual simulator track?

[vguerra]

yes.. i can remove the predictive steering. But probably the parameters would need to be tuned.