clearpilot/sessions/2026-04-26-1229-hyundai-canfd-steer-torque-bump/README.md

# Session: 2026-04-26 — Hyundai CAN-FD steering torque bump

Documentation for commit `4058269` on `clearpilot`. Reference for any
future re-tuning.

## What changed

Bumped this Tucson HDA2 (CAN-FD platform) from comma's conservative
steer ceiling to a value that gives more headroom on tighter on-ramp
clovers, plus a small rate-limit nudge so the controller can actually
reach the new ceiling within reasonable transient time.

| Constant | Before (comma default for CAN-FD) | After (this commit) | Comma's non-CAN-FD HKG default |
|---|---:|---:|---:|
| `max_steer` / `STEER_MAX` | 270 | **324** | 384 |
| `max_rate_up` / `STEER_DELTA_UP` | 2 | **3** | 3 |
| `max_rate_down` / `STEER_DELTA_DOWN` | 3 | **5** | 7 |
| `max_rt_delta` | 112 | **134** | ~150 |

## Where the change lives — TWO files in lockstep

The panda safety firmware enforces these limits **independently** of
openpilot. If only one side is bumped, panda rejects the larger
commands and you get cut-out / `commIssue` behavior. **Always change
both, always to the same numbers, in the same commit.**

1. **`panda/board/safety/safety_hyundai_canfd.h`** (lines 3-19)
   - `HYUNDAI_CANFD_STEERING_LIMITS` struct: `max_steer`, `max_rate_up`,
     `max_rate_down`, `max_rt_delta`.
   - This is C; modifying it changes the panda firmware hash, which
     forces an automatic re-flash on next `pandad` start. No manual
     panda flash command needed.

2. **`selfdrive/car/hyundai/values.py`** (CAN-FD branch of
   `CarControllerParams.__init__`, lines 29-36)
   - `STEER_MAX`, `STEER_DELTA_UP`, `STEER_DELTA_DOWN`.
   - Pure Python; picked up on next `controlsd` start.

## What each constant does

- **`max_steer` / `STEER_MAX`** — peak torque magnitude the controller
  can request. Hard ceiling. Going past this is the headline "more
  torque" knob.
- **`max_rate_up` / `STEER_DELTA_UP`** — per-100Hz-cycle upward slew
  cap. Higher = faster ramp into a turn. With `max_rate_up = 3` and
  `max_steer = 324`, time from 0 to ceiling is 324 / 3 = 108 cycles =
  1.08 s.
- **`max_rate_down` / `STEER_DELTA_DOWN`** — per-cycle downward slew
  cap. Higher = faster release back toward straight. We chose 5 (vs
  comma's 7) for a smoother release feel.
- **`max_rt_delta`** — cumulative torque change allowed across a
  rolling 250 ms window (`max_rt_interval`). It's a long-window
  envelope check, separate from the per-cycle rate. Should scale with
  `max_steer` — we used `max_steer × ~0.41` to mirror comma's ratio.
- **`driver_torque_allowance` / `STEER_DRIVER_ALLOWANCE`** — driver
  wheel torque (Nm read off the wheel) that's tolerated before the
  system starts derating its own command. Left at 250.
- **`driver_torque_factor` / `STEER_DRIVER_MULTIPLIER`** — how
  aggressively the system fights driver input above the allowance.
  Left at 2.

## How to go higher (path to 384)

`384` is the community-consensus safe ceiling for HKG. Comma uses it
as the default for every non-CAN-FD HKG that isn't on the explicit
255-blacklist, and they merged it for HDA1 CAN-FD (EV6 / Ioniq 5) in
[openpilot PR #25723](https://github.com/commaai/openpilot/pull/25723).
The PR author noted "max steer needed to be 384 to make basic turns."
Tucson NX4 HDA2 is **not** on the 255-blacklist (see
[issue #24122](https://github.com/commaai/openpilot/issues/24122) for
the verified blacklist).

To go from 324 → 384:

```c
// panda/board/safety/safety_hyundai_canfd.h
.max_steer = 384,
.max_rt_delta = 158,         // ~max_steer × 0.41
.max_rate_up = 3,
.max_rate_down = 5,          // or 7 for comma-matched aggressive release
```

```python
# selfdrive/car/hyundai/values.py CAN-FD branch
self.STEER_MAX = 384
self.STEER_DELTA_UP = 3
self.STEER_DELTA_DOWN = 5    # or 7
```

Beyond 384 is uncharted for HKG — comma has not tested past it. Some
forks (sunnypilot has discussions) try higher for very heavy vehicles
but with mixed results. Don't go past 384 without an explicit reason.

## Things to watch for / community-flagged risks at higher values

1. **EPS time-out cut every ~90 frames.** The CAN-FD safety already
   forces a brief torque cut to stop the EPS from faulting (the
   `min_valid_request_frames = 89`, `max_invalid_request_frames = 2`,
   `min_valid_request_rt_interval = 810000` block of
   `HYUNDAI_CANFD_STEERING_LIMITS`). This is independent of
   `max_steer`. Bumping the ceiling does not lengthen the cut-free
   window — you just hold the higher torque for the same ~890 ms before
   the brief cut. If you're getting `Steering Temporarily Unavailable`
   *during sustained turns*, the issue is this cut, not the ceiling,
   and it can't be tuned without risking a real EPS fault.

2. **`steerTempUnavailable` / "Cruise Fault: Restart the Car".** Has
   been reported on cars in the 255-blacklist when pushed to 384.
   Tucson NX4 is not blacklisted, so 324–384 is normally safe — but if
   you see this alert during slow sweeping turns, that's the symptom.
   Roll back to 270 and confirm.

3. **Lateral accel retune.** Higher torque headroom can cause the
   torque controller to overshoot if `latAccelFactor` was tuned for
   the old ceiling. EV6/Ioniq 5 testing in PR #25723 had to drop
   lateral accel to 2.5 m/s² when bumping from 270 to 384. Watch for
   over-correction (zig-zag in lane center) after a bump and retune
   `latAccelFactor` in `selfdrive/locationd/torqued.py` or via the
   torque-tune Params if needed.

4. **Driver-fight feel.** `driver_torque_allowance = 250` /
   `driver_torque_factor = 2` is the blending knob. Most "openpilot
   fights my hands" complaints come from people who lowered allowance
   or raised factor. Don't touch these without a specific reason.

5. **Panda safety hash mismatch.** Changing
   `safety_hyundai_canfd.h` regenerates the panda firmware binary with
   a different signed hash. On the next `pandad` start, pandad detects
   the mismatch and re-flashes the panda automatically (see
   `pandad.log`: "Panda firmware out of date" → "flash: flashing" →
   "Done flashing"). Takes ~10 s. No manual action needed; just expect
   a brief delay before controls come up.

## Verifying the change took effect

```bash
# 1. Confirm the rebuild happened and panda firmware was re-signed.
grep "panda/board/obj/panda_h7" /tmp/build_only.log 2>/dev/null   # or run build_only.sh and watch for "signing N bytes"

# 2. Watch for re-flash on launch.
tail -f /data/log2/current/pandad.log
# Should see: "Panda firmware out of date, update required"
#             "flash: flashing"
#             "Done flashing"

# 3. Confirm the openpilot side is using the new value.
su - comma -c 'PYTHONPATH=/data/openpilot python3 -c "
from cereal import car
from openpilot.common.params import Params
cp_bytes = Params().get(\"CarParams\")
with car.CarParams.from_bytes(cp_bytes) as cp:
    print(\"safetyConfig safetyParam:\", [c.safetyParam for c in cp.safetyConfigs])
"'

# 4. Live-check the actual commanded torque ceiling (drive a moderate turn).
# carControl.actuators.steer should now be able to peak above 0.79 (270/255 normalized)
# but stay under 1.0 (full saturation at the new ceiling).
```

## Rollback

If anything misbehaves, revert just the two-file commit:

```bash
git revert 4058269
chown -R comma:comma /data/openpilot
su - comma -c "bash /data/openpilot/build_only.sh"
# Next pandad launch will re-flash back to 270.
```

## Sources

- [openpilot PR #25723 — HDA1 EV6/Ioniq 5 270 → 384](https://github.com/commaai/openpilot/pull/25723)
- [openpilot issue #24122 — HKG torque blacklist verification](https://github.com/commaai/openpilot/issues/24122)
- [openpilot PR #26427 — Hyundai Tucson 2023 support](https://github.com/commaai/openpilot/pull/26427)
- [sunnypilot — increasing torque help](https://community.sunnypilot.ai/t/increasing-torque-help-needed/2082)
- [sunnypilot — raising torque for heavier vehicles](https://community.sunnypilot.ai/t/raising-the-torque-for-heavier-vehicles/862)