Skip to content

Rival passing and emergency traffic-braking reconciliation — source QA

Status: integrated actual Unreal AI source, not a UE5.8 engine-compiled or installed playable build. The 3D car remains a simplified Chaos rigid-body prototype without independently simulated wheel contacts.

Problem confirmed in the source

PR #287 made the only genuine AI rival steer around a slower player, but allowed passing to begin 900–1600 cm behind the player even when the measured closing speed was very high. Meanwhile the existing independent YieldRivalToPlayer emergency stop uses true forward closing velocity and a 380 cm safety gap. For example, a car 1050 cm behind and closing at 1350 cm/s has a predicted time to the safety margin of about 0.50 seconds; the real traffic policy correctly brakes immediately, even though the car's passing controller may have just ordered a sharp lane-change steer. The previous integration retained that steer while braking. That is an unsafe control conflict.

Actual new behaviour

The existing physical ALatebrakeRivalCar::Tick now:

  • Allows new passing only on the two longer 5600 cm straight approaches of the original four-gate circuit, never the shorter 3600 cm straights where road length is insufficient for this untested manoeuvre.
  • Requires the slower player to be 1600–2400 cm ahead, relative closing speed 300–900 cm/s, time to the conservative contact margin of at least 1.8 seconds and at least 4200 cm before the approaching corner. If these checks fail, the original traffic governor still slows/brakes the rival normally.
  • Keeps the original physical passing-steer command within a bounded road corridor, with no new engine power or modified gate position.
  • Runs the original independent real-world-velocity traffic safety policy after selecting a passing steer, followed by its existing 0.18-second following-brake hold.
  • Calls ResolveRivalPassTraffic with the original route command, active passing result and traffic-held command. If the real traffic system brakes, the pass is cancelled and steering returns to the original route direction while the brake remains engaged. The passing state resets so it cannot silently reappear mid-brake.
  • Preserves existing rival yaw/spin moderation, verified checkpoint crossings, recovery placement, race times, reward and saved-profile authority, without touching the player's controls.

The correction does not disable emergency braking to make passes easier. It avoids beginning a manoeuvre too close to impact, and fails safe if live traffic still requires braking.

Automated qualification

  • Tests/Standalone/RivalStraightPassAssistTests.cpp now rejects the old unsafe short-distance and high-relative-speed starts; validates safe long-straight entries, shorter-straight refusal, both lane choices, lane-change abort and 20,000 sampled road states.
  • Tests/Standalone/RivalPassTrafficResolutionTests.cpp checks physical same-lane emergency braking, actual separated-lane passing, normal prior braking, existing brake-hold hysteresis, malformed control inputs and 20,000 diverse traffic encounters.
  • Tests/Editor/UnrealRivalPassTrafficResolutionBindingTests.py verifies the actual AI pawn calls the new reconciliation between real traffic/hold decisions and yaw/spin/physical motor output.
  • All portable tests run strict C++20 on Windows, Linux and macOS; source-binding tests do not compile Unreal.

Still required for playable racing qualification

  1. Compile actual Unreal 5.8/UHT and cook the original map on an authorized Windows workstation. Install the real 3D candidate.
  2. On both long straights place a slow player 16–24 m in front and record actual passing steering, brake demand, separation, speed, turning radius and verified gate progression.
  3. Repeat with high closing speed, player sudden braking, player swerving into pass lane, overlapping cars and a traffic-brake hysteresis already active; confirm no aggressive passing steer persists under emergency brake.
  4. Confirm no pass on the shorter straights or near corners; test road boundaries, vehicle collision shapes and 30/60/120 FPS.
  5. Test restart/focus loss/recovery/invalid sample and compare winning/finishing only from actual directional gate crossings.
  6. Improve after real physics telemetry—not based on synthetic portable tests alone.

Playable-release qualification remains tracked in issue #148.