Allow release recovery only when fresh measured yaw undertracks both aligned current and delayed requests and PSCM limit is below 2. Unwind the opposing bias toward zero using current yaw error and existing antiwindup; preserve C0, base geometry, gains, rates, and safety guards.
Add mirrored unit checks and a sanitized recorded turn-exit regression. Validate with 142 tests and 97 subtests, full-route frozen-input replay, large-turn retention, CAN packing, diagnostics, and generated Sunnylink schema checks. Physical improvement remains unvalidated.
Reuse the existing model-pose allocator for aligned large maneuvers while encoding remaining selected curvature as C0/C1 and keeping C2/C3 zero. Permit measured heading backoff during release or PSCM limits without turning model-base changes into stored bias.
Validate with 127 tests and 67 subtests, including recorded large-turn retention, release and reversal, repeated-measurement backoff, CAN packing, logging, and Sunnylink schema checks. Replay checks command behavior; enabled vehicle tracking remains unvalidated.
Retain the absolute desired-curvature base and unchanged C0, while adding
measured yaw-error correction to C1 under fresh PSCM status. Preserve command
limits, reset on override or unusable status, and release stored correction
with the base request. Admit reachable partial increments at host slew limits.
Publish PSCM enums with original CAN receipt timestamps through carStateSP.
Add telemetry, status/driver guards, CAN roundtrip tests and recorded fixtures.
Validation: focused suite 116 tests and 38 subtests; Ruff, settings compilation
and diff checks pass. Production replay covers 52,273 route80 cycles; no-status
fallback preserves v4 over 246,961 cycles / 43 segments. Physical stability and
the reported 85-degree plateau remain unvalidated; EPS limits can inhibit the
new correction.
Derive full absolute C1 heading from the same selected curvature as C0, retaining existing bounds and independent slew. Keep the former filtered model heading as a diagnostic comparison and preserve input validity gates.
Add real route80 command regressions and release/reversal checks. All 97 focused tests pass; 43-segment replay preserves C0 and gates exactly and matches the independent C1 candidate. Physical tracking and stability remain unvalidated for this revision.
Encode the selected bounded curvature as C0 with an 8 m minimum preview while retaining full model-heading C1. Limit each channel independently so C1 transitions cannot delay C0 release, and validate the selected action source timestamp.
Add action, release, source-freshness, CAN and route regressions. Document the slow-turn reference disagreement and the limits of frozen-motion replay; physical centering remains unvalidated.
Replace the weak nominal acceleration conversion with C2-free spatial path requests. Align retained model geometry using measured CAN yaw before filtering model innovations, and preserve large-turn demand and straight-path centering.
Validate with seven recorded maneuver episodes, full-route command replay, real CAN packing and focused controller/settings tests. Physical closed-loop behavior remains unvalidated.
Track the bounded planner reference through C0 and delay-aware PI/rate feedback through C1. Remove the failed Shared Path toggle and add a default-off, Lightning RL38-specific Virtual Angle setting. Reject stale inputs and disable outgoing lateral requests when the path is invalid.
Validation: 85 focused tests plus 22 subtests, native Params, real CAN packing, settings generation and Ruff passed. Frozen route78 replay attenuates the observed command forcing; physical stability and turn authority remain unvalidated.
Preserve geometric C0/C1 before nominal plateaus during same-direction buildup. Keep reversal guards, command limits, C2 policy, and the existing default-off Shared Path Controller selection. Require nonzero demand for joint buildup.
Known limitation: nominal short-turn cancellation settles later with queued commands. Retain that regression as an explicit expected failure; this experiment does not establish physical response or resolve unwind. Add entry and zero-demand coverage.
Assisted-by: OpenAI Codex
Continue the same allocator through a short turn and cancellation so release checks retain command lead as well as nominal coefficient state. Reject earlier geometry buildup that keeps charging after cancellation. No production controller changes.
Assisted-by: OpenAI Codex
Update opendbc to 72a775d3 for C0/C1/C3 wire-range saturation and nonfinite input rejection. No fallback controller or tuning changes are included.
Assisted-by: OpenAI Codex
Retain larger model-derived fast fields when nominal allocation is equivalent, with per-field plateau qualification and inward-demand release priority. Keep corrected and geometric fast fields independently selectable without changing the existing contribution map, C2 policy, cadence, or command limits.
Validated with 68 controller/fallback/logging tests, 5 adversarial release tests, and 38300 fixed-input replay updates. Physical turn authority and release remain unverified; the existing experiment stays default off.
Assisted-by: OpenAI Codex
Cache per-field packet conversion and state projections within each allocation instead of recomputing them for every candidate combination. Preserve candidate ordering, scores, limits, and selected commands. Add a deterministic limiter-work regression budget.
Local recorded-input mean controller CPU time falls 56%; 1292 recorded updates and 4000 randomized allocations match the previous outputs exactly. Device timing remains unverified.
Assisted-by: Codex
Use SwagLogger.event for controller selection and periodic diagnostics. Logger.info forwards arbitrary keywords to Logger._log and crashed all Ford startups, regardless of the experiment toggle. Exercise both actual call sites with INFO enabled and the real logger/formatter.
Assisted-by: Codex
Separate holding demand, bounded pose feedback, and nominal coefficient allocation. Add a default-off Sunnylink selector with startup diagnostics and preserve the existing controller when disabled.
Assisted-by: Codex
Remove model-pose residuals and tracking trim from the gentle regime. Blend the model pose into C0/C1 only as maneuver demand rises, while retaining opposing-path C2 unload and the coordinated 100 Hz handoff.
Assisted-by: Codex
Use constrained desired curvature for ordinary C2 while keeping model geometry authoritative in the coordinated C0/C1 residual. This restores normal centering strength without changing large-maneuver or bounded-feedback behavior.
Assisted-by: Codex
Keep the model's remaining path as feedforward while using the delay-aligned measured pose only as a bounded trim. Allocate common gentle model curvature to C2 and carry changing geometry in C0/C1 without allowing the action head to invent a path.
Assisted-by: Codex
Rebase model preview against a gainless 100 ms curvature-trend prediction. Remove direct local-curvature feedback while preserving coordinated C0/C1/C2 authority and geometric C2 unloads.
Assisted-by: Codex
Use the first two meters of model heading for measured-curvature feedback while preserving the existing longer model-pose feedforward. This prevents future geometry from initiating premature correction without weakening turn anticipation.
Assisted-by: Codex
Use measured curvature error against the forward model path to add bounded bidirectional C0/C1 correction. Unload stale C2 when it would oppose an unwind or reversal.
Assisted-by: Codex
Return to the pre-predicted-pose C2-first controller from a1dcec490 after road testing found both later variants weaker or unstable. Preserve the current sunnypilot master merge and 100 Hz LMC2 transport.
Assisted-by: Codex
Remove delay-projected measured-curvature feedback that amplified curve hunting. Keep the model polynomial in the current vehicle frame while preserving the coordinated 100 Hz C0/C1/C2 handoff.
Assisted-by: Codex
Advance the rolling model path by the generic lateral delay, express its remaining seven-meter pose in the predicted vehicle frame, and derive C2 from the same steady geometry. Remove desiredCurvature as a competing Ford path target.
Assisted-by: Codex
Move heading authority lost at the DBC angle limit into available C0 endpoint authority while retaining the coordinated output limiter.
Assisted-by: Codex
Use the model's forward offset and heading for fast path authority while C2 retains ordinary path following. Coordinate all transmitted coefficients through one bounded handoff and add measured-curvature catch-up without overshoot countersteer.\n\nAssisted-by: Codex
Use heading/action hysteresis to keep normal driving entirely on C2 and large maneuvers entirely on model C0/C1. Restore a one-second pose horizon with a 7 m floor.
Assisted-by: Codex
Keep desired curvature in C2 for ordinary driving, then continuously hand off to model offset and heading for large maneuvers. Use one fitted 0.5 second lookahead with a 7 meter floor for both pose fields and keep C3 zero.
Assisted-by: Codex
Remove the one-frame C2 persistence and restore the continuous gentle-centering allocation. Real turn demand now clears C2 immediately and remains in the bounded fast path fields.
Assisted-by: Codex
Delay C2 by one 50 ms model frame and place the new-request difference in C0/C1 alongside measured tracking error. Clear the delay on inactive or invalid control.
Assisted-by: Codex
Keep upstream-style desired curvature active in C2 for steady path following. Use C0/C1 only for demand beyond C2 and measured tracking error, preserving fast turn and unwind authority without replacing C2.
Assisted-by: Codex
Make C0 and C1 a coherent virtual-curvature pair sourced from the constrained action target and measured tracking error. Keep model trend only for supplemental C2 unloading so model geometry cannot inflate fast steering authority across vehicles.
Assisted-by: Codex
Road testing showed the endpoint-constrained C0/C1 pair opposed the requested rotation and delivered less than half the needed authority. Restore the prior same-direction fast-path encoder.
Assisted-by: Codex
Isaac Barham