HKG: Car Port for EV6 2025-26 (HDA2, LFA2)

[sunnyhaibin]

Based on #1646

Juan's 2025 EV6 Wind (HDA2): 3fc2f65852a45c13

[sshane]

If we used EV6 params for everything, all other cars would violate accel and jerk ISO limits:

Set to Sportage, most conservative car:

2 m/s^2 is not terrible, but not ideal, especially if we port a new car that bumps everything down further. Need some way in panda safety to differentiate the cars

@elkoled can you run the script for Subarus to see the worst case and change to most conservative car?

[devtekve]

Need some way in panda safety to differentiate the cars

I was working a while ago on intercepting the fingerprint by tapping to the UDS messages, I didn't love where it was heading and thought that it would just consume "a lot" of memory on the red panda for storing each fingerprint needed...

If you are curious, this is the code I was writing back then https://github.com/sunnypilot/opendbc/pull/294/changes

[sshane]

Trying a kalman filter instead of low pass filter, overshoots, but interesting to try:

[sshane]

@sunnyhaibin please try this latest tune out on your vehicle (not the commits after this, I haven't tested them yet!)

[sshane]

Stock route with lots of turns and other scenarios: 3fc2f65852a45c13/00000043--0faaacc68d

[devtekve]

@sshane why not EMA like we have on the SP branch tho?

https://github.com/sunnypilot/opendbc/blob/1190d6672f71670c43795a9ed5c42c4803adf21d/opendbc/car/hyundai/carcontroller.py#L73-L75

[sshane]

@sshane why not EMA like we have on the SP branch tho?

https://github.com/sunnypilot/opendbc/blob/1190d6672f71670c43795a9ed5c42c4803adf21d/opendbc/car/hyundai/carcontroller.py#L73-L75

What do you mean? FirstOrderFilter class is a low pass filter/ema. The gains on that branch are pretty close to what I found, but lessens rc earlier.

Converted your alphas to rc:

  SMOOTHING_ANGLE_VEGO_MATRIX = [0, 8.5, 11, 13.8, 18]
  SMOOTHING_ANGLE_RC_MATRIX = [0.19, 0.09, 0.0233, 0.0067, 0.0]

[devtekve]

@sshane why not EMA like we have on the SP branch tho?
https://github.com/sunnypilot/opendbc/blob/1190d6672f71670c43795a9ed5c42c4803adf21d/opendbc/car/hyundai/carcontroller.py#L73-L75

What do you mean? FirstOrderFilter class is a low pass filter/ema. The gains on that branch are pretty close to what I found, but lessens rc earlier.

Converted your alphas to rc:

  SMOOTHING_ANGLE_VEGO_MATRIX = [0, 8.5, 11, 13.8, 18]
  SMOOTHING_ANGLE_RC_MATRIX = [0.19, 0.09, 0.0233, 0.0067, 0.0]

Thanks for the explanation! I didn't realize you had reverted the kalman filter 😅
I'll try with the FirstOrderFilter class with your values and report back.

The values I have on our branch are from a lot of trial and error, aiming to drop the smoothing as soon as possible, hence why the RC lessens earlier

---

I've also made a bit of tweaking on the torque reduction gain calculation to more dynamically adjust given that EU and US markets are quite different, so I generally have lower torque reduction gain values but with angle error taken into account to boost the torque dynamically.

However, my changes are based what claude did on the hackathon, so i will actually pull your changes and test those too.

this is what we had for the reduction gain https://github.com/sunnypilot/opendbc/blob/1190d6672f71670c43795a9ed5c42c4803adf21d/opendbc/car/hyundai/carcontroller.py#L37-L50

[devtekve]

@sshane i could daily drive with your tuning! It's working pretty well, great work!

This is the route 4dcb6256482436c6/000001f5--382c61ee17

The only "caveat" I have is that it felt too stiff to override at lower speeds, I think it's possible the torque reduction gain starts too high ? But otherwise no jitter on the path, no EPS noises that I noticed and no understeer or oversteer.

[sshane]

@sshane i could daily drive with your tuning! It's working pretty well, great work!

This is the route 4dcb6256482436c6/000001f5--382c61ee17

The only "caveat" I have is that it felt too stiff to override at lower speeds, I think it's possible the torque reduction gain starts too high ? But otherwise no jitter on the path, no EPS noises that I noticed and no understeer or oversteer.

which commit?

[sshane]

Here's a video showing visually how second order filter smooths out the stair steppy model requests a bit better:

https://connect.comma.ai/3fc2f65852a45c13/00000062--4aa5f8800c

Peek.2026-04-08.23-53.mp4

[devtekve]

@sshane i could daily drive with your tuning! It's working pretty well, great work!
This is the route 4dcb6256482436c6/000001f5--382c61ee17
The only "caveat" I have is that it felt too stiff to override at lower speeds, I think it's possible the torque reduction gain starts too high ? But otherwise no jitter on the path, no EPS noises that I noticed and no understeer or oversteer.

which commit?

I was on my own branch to use ioniq 5 PE limits and with the safety fix from #3294 applied, but this was the commit
sunnypilot@c72b6fd

---

Edit: to be clear, I tested the changes you had before you disabled the filtering @ 99664ba ;