mirror of
https://github.com/sunnypilot/sunnypilot.git
synced 2026-11-12 11:23:43 +08:00
Isaac Barham
3d6053945a
Routes 93, 94 and 96 (CD210, f6a08afbf) had no crashes, model drops or
params errors, but turns were unreliable:
- The turn-in hold cut a turn already under way. The cue came only once
the model's turn prediction reached 0.35, after the car had turned
~80 deg on its own; the hold read the corner as just reached and held
1.5 m, the wheel unwound and the driver took over (route 93, 58.6 s).
- The model drops the turn 6.5-6.6 s after the cue (routes 92-96): the
desire is a one-frame pulse and its history is ~5 s. After a cue at a
red light the turn was gone by pull-away, and the turn-state gate then
switched the floor off too, so the car went straight (route 96, 8 s).
- The hold used up that window: released at 705.3 s on route 96 as the
model's turn state fell from 1.0 to 0.19.
Turn steering is now only the first turn plan floor: the command never
falls below 0.75x the model's plan toward the turn, at every turn speed,
and does nothing when the plan points out of the turn. Removed: the
hold and its TurnInHold toggle (params library back to the PlanLong
build that ran on the car), the look-ahead, the route scaling, the
4-8 m/s fade and the turn-state gate. The corner speed (openpilot
longitudinal only) and navd finding the turn for any model stay. The
shaping log fields in modelDataV2SP are deprecated in place.
Without the gate, the early turn-in pulling away from a stop that it was
added for (routes 85/89) can come back. Not yet driven.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Docs development
The docs/ tree is the source for docs.comma.ai.
The site is updated on pushes to master by this workflow.
1. Build the site
op docs --build
2. Run the site locally (rebuilds on change)
op docs