Blend MPC back in over the final 5 mph of the CEM limit while preserving strong E2E braking. Slew positive acceleration and freeze the integrator during the experimental-mode exit so the handoff stays smooth. Keep the existing lead and confidence gates on the release path.
Original PR: #115 by @1454
Force Stop is now better at faster approach speeds. It has more authority on braking in general.
Force Stop Distance Offset: still the lever that moves the ending point sooner or later.
Radar bug fixed
Keep the Bosch-A lead probability filter cadence change limited to the verified Honda Bosch-A path. Preserve the existing radar timestep for every other platform and add regression coverage for the gate and both timing paths.
Co-authored-by: JamesL787 <jamezelle1@gmail.com>
is_bosch_a_radar_car() and its downstream handling in radard.py never
landed alongside the radar decoder (interface.py/values.py/radar_interface.py)
-- without it, the newly-decoded Bosch-A radar points feed into radard's
generic path, which assumes the wrong Kalman filter timestep and is missing
a real staleness/freshness gate this radar's actual cadence needs.
Ports (unmodified in behavior, only renamed to match this repo's
HONDA_BOSCH_A / HONDA_BOSCH_A_RADAR_VERIFIED naming instead of "civic_bosch"):
- Kalman filter dt selected from the radar's real ~15Hz cadence
(HONDA_BOSCH_A_RADAR_TS), not the ~20Hz model-cycle default
- measurement_update gating so a duplicate liveTracks payload (radar runs
slower than the model loop) isn't absorbed twice by the KF
- a "preferred track" continuity path: a previously-selected lead stays
selected through ordinary model-probability fluctuations as long as it's
still coherent, instead of re-arbitrating from scratch every cycle
- staleness detection with two independent triggers (a better-scoring
challenger track sustained over several cycles, or gross absolute-range
disagreement with the model sustained over several cycles) that drops a
stale preferred track rather than holding it indefinitely
- honda_bosch_a_low_speed_radar_lead_sane(), requiring a few real sweeps
before a radar-only low-speed takeover
Tested on the branch this was developed on (109/109 relevant tests pass);
this port is a mechanical, unmodified transplant onto this repo's current
radard.py -- worth a run through this repo's own test suite before merge.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
(cherry picked from commit 8d23733946)
Avoid EPS torque faults when entering always-on lateral while the driver is steering. The behavior is scoped to HYUNDAI_KONA_NON_SCC and leaves other vehicles on the shared lateral-active path.
Original contribution by firestar5683.
firestar5683