kalman VOACC

[sshane]

WIP. dumb kf on x is not good, because we don't use xStd? or use ego at all?

(only dRel is relevant)

Using xStd works pretty well!

(only dRel is relevant)

• doesn't handle lead changes since model's output always smoothly moves around

[sshane]

Here the model predicted the lead to be going above 25 m/s when it first detected it, which was way off, causing no braking whatsoever: https://connect.comma.ai/0f79c454f812791a/000000a3--e1e54de187/1037/1067

Using a kf to predict lead velocity from distance + compensating with aEgo, it instantly snaps to ~15 m/s which is what I roughly calculated with derivative on the model lead distance. Plus my speed at the lowest point was 17 m/s and I was still faster than the lead, so the kf's prediction of 15 m/s instantly upon detecting the lead looks much better!

---

Left is simple derivative of distance over time, right is kf with xStd and aEgo compensation (so kf doesn't have to deal with dRel changes from ego)

[morrislee]

Here the model predicted the lead to be going above 25 m/s when it first detected it, which was way off, causing no braking whatsoever: https://connect.comma.ai/0f79c454f812791a/000000a3--e1e54de187/1037/1067

Using a kf to predict lead velocity from distance + compensating with aEgo, it instantly snaps to ~15 m/s which is what I roughly calculated with derivative on the model lead distance. Plus my speed at the lowest point was 17 m/s and I was still faster than the lead, so the kf's prediction of 15 m/s instantly upon detecting the lead looks much better!

Left is simple derivative of distance over time, right is kf with xStd and aEgo compensation (so kf doesn't have to deal with dRel changes from ego)

Is this why coming to a lead car on the highway takes a while to converge to the lead car speed? (Usually gets close and brakes hard, oscillating behind it a few times to match lead car speed)

[sshane]

yes. @haraschax is working on a new model to fix this