Traffic.h

叠加在 ORCA 避障之上的交通规则共享 fragment FTrafficShared(FA_MassBattleBaseSharedFragment 派生)、枚举 ETrafficSide 与 8 个规则子结构。每条规则有独立开关且默认全开;总闸 FAvoidance::bUseTrafficRules(默认关)是启用整个交通层的唯一开关,每条规则 bEnable 关闭时干净退化。规则分三类:预解算机动原语(求解器运行前施加于 DV:SideStep 仅静止,LaneShift 与 Flank 仅移动,WallSlide 始终生效)、责任仲裁(ComputeAvoidanceRatio 内逐对应用:单一 Arbitration 让步矩阵 + Combat.RightOfWay 攻击者路权 + Arbitration.ReturnPriorityDist 远归位路权)、求解器范围规则(LivingVsDying 采集过滤 / Combat.SkipAvoidance 整段跳过)。质量加权由全局唯一的 Arbitration.MassWeighting 统一承担(取代旧各规则独立的 MassWeightingStrength)。以可变共享片段分发:相同 authored 值在每个原型中合并为一个实例(模拟 FTrafficShared / FAnimShared 模式),同时允许 GetMutableSharedFragment 运行时调整。

Traffic-rule shared fragment FTrafficShared (derived from FA_MassBattleBaseSharedFragment) layered on top of ORCA avoidance, plus the ETrafficSide enum and 8 rule sub-structs. Each rule has an independent on/off switch and all default ON; the master gate FAvoidance::bUseTrafficRules (default OFF) is the single switch that arms the whole layer, and each rule's bEnable degrades cleanly when off. The rules fall into three classes: pre-solve maneuver primitives (applied to DV before the solver runs: SideStep is idle-only, LaneShift and Flank are moving-only, WallSlide is always on), yield arbitration (applied per neighbor pair inside ComputeAvoidanceRatio: the single Arbitration yield matrix + Combat.RightOfWay attacker right of way + Arbitration.ReturnPriorityDist far-returner right of way), and solver-scope rules (LivingVsDying gather filter / Combat.SkipAvoidance whole-solver bypass). Mass weighting is carried by the single global Arbitration.MassWeighting (replacing the former per-rule MassWeightingStrength knobs). Distributed as a mutable shared fragment so identical authored values dedupe to one instance per archetype (mirroring the FTrafficShared / FAnimShared pattern) while still allowing runtime tweaks via GetMutableSharedFragment.

对头车道形成的世界侧约定,由 FTrafficArbitration::TrafficSide 使用;歧义判侧(横向证据落在 CommitBand 死区内时)也退回该约定。

World-side convention for head-on lane formation, used by FTrafficArbitration::TrafficSide; ambiguous side picks (lateral evidence inside the CommitBand dead-zone) also fall back to it.

采集阶段过滤规则:关闭时,死亡邻居从 RVO 邻居列表中移除——自身可穿过倒下身体。PBD 物理碰撞由 Collider.bCollideWithDyingAgent 单独控制。

Gather-phase filter rule: when disabled, dying neighbors are removed from the RVO neighbor list — self walks through falling bodies. PBD physical collision is governed separately by Collider.bCollideWithDyingAgent.

L2 责任仲裁(Docs/TrafficRedesignPlan.md):单一让步矩阵取代旧五条责任规则(ArrivedVsReaching / OffPathVsOnPath / NonObstAvoidingVsObstAvoiding / NonAttackerVsAttacker / AttackerVsAttacker)。逐对裁决谁让多少:档位低方、离路径方、同向追尾后方、封锁面对方统一承担 YieldBias,对方承担其余;唯一全局 MassWeighting 缩放不对称对(取代旧四条规则各自的强度)。同时携带机动层共享参数:行进侧约定(TrafficSide)、判侧死区(CommitBand 滞回)、统一幅度语义(最大速度 × BiasSpeedFrac,BiasCapFactor 封顶)。关闭 → 旧基线(可让对称 0.5、不可移动 1.0)。

L2 yield arbitration (Docs/TrafficRedesignPlan.md): a single matrix replacing the five former responsibility rules (ArrivedVsReaching / OffPathVsOnPath / NonObstAvoidingVsObstAvoiding / NonAttackerVsAttacker / AttackerVsAttacker). ComputeYield decides per pair who yields how much: the lower-rank side, the off-path mover, the same-lane rear and the blockade-facing agent take YieldBias, the other side the rest; one global MassWeighting scales asymmetric pairs (replacing the four per-rule strengths). Also carries the shared maneuver-layer knobs: the traffic-side convention (TrafficSide), the side-pick dead-zone (CommitBand hysteresis) and the unified magnitude semantics (MaxSpeed × BiasSpeedFrac, capped at × BiasCapFactor). Disabled → legacy baseline (symmetric 0.5, immovable 1.0).

战斗与交通的交互(合并旧攻击者路权与 AttackerSkipAvoidance 两条规则):攻击状态单位与避障/交通系统的交互方式,两个独立子开关。状态模型:Aim/前摇/后摇与追击接受半径内的 Cooling 是同一个锁定态(站定、有路权);超出追击接受半径的 Cooling 是攻击追击——Agent 主动走近目标并正常参避(Attack.Range > Chase.AcceptanceRadius 的压进配置:远射程开火、冷却期间压进)。

Combat interplay with traffic (merges the former attacker right-of-way and AttackerSkipAvoidance rules): how attack-state agents interact with the avoidance/traffic system, via two independent sub-toggles. State model: Aim/PreCast/PostCast and Cooling-within-chase-acceptance are ONE locked state (planted, right of way); Cooling BEYOND the chase acceptance is attack-chase — the agent actively walks in and participates in avoidance (the Attack.Range > Chase.AcceptanceRadius press-in config: fire from max range, close in during cooldown).

绕行机动原语:移动非攻击者旋转 DV 以绕开前进路径走廊内的挡路者——不可移动体(bCanAvoid=false)与更重单位一并处理(旧两种模式合并,由责任仲裁决定谁绕)。带缝隙瞄准:走廊内 ≥2 个阻挡者夹出横向缝隙且净宽(边到边)≥ 2×自身半径 + MinGapWidth 时,以缺口缩放的前馈偏置瞄准缝心穿过(对准即归零);更窄的缝物理上不可通行,直接排除(窄缝扑口是堆积主因之一)。无可钻缝 → 逐阻挡者距离加权远离推力(绕开整簇)。所有侧向量均在路径绝对系下计算。

Flank maneuver primitive: a moving non-attacker rotates its DV to clear blockers in its forward path corridor — immovable bodies (bCanAvoid=false) and heavier agents alike (the two former Flank modes merged; the yield matrix already arbitrates who detours). Gap seeking: when ≥2 corridor blockers bound a lateral gap whose free width (edge-to-edge) ≥ 2×SelfRadius + MinGapWidth, a deficit-scaled feedforward nudge aims at the gap center (magnitude → 0 as self lines up); narrower gaps are physically impassable and excluded (mouth pile-up). No threadable gap → per-blocker distance-weighted away-nudge (flank around the cluster). All lateral quantities are computed in path-absolute coordinates.

车道偏移机动原语(取代 HeadOnLaneBias):移动自身向来车反方向、行进侧约定侧旋转 DV。对头锥为结构阈值(≈135°,不可调);幅度按责任加权——对称对保持历史每邻居角度,不对称对倾斜。

Lane-shift maneuver primitive (replaces HeadOnLaneBias): moving self rotates its DV away from oncoming moving neighbors, toward the TrafficSide convention. The head-on cone is a structural threshold (≈135°, no tuning knob); magnitude is yield-weighted — symmetric pairs keep the historical per-neighbor rotation, asymmetric pairs tilt.

侧让机动原语(取代 OncomingSideStep):仅静止自身沿每个来袭邻居速度的垂直方向添加横向 DV 分量以侧让,而非被推着走。归位者与移动者同属在途,走 LaneShift 而非侧让(在途有真实期望速度,车道旋转比逐邻居侧让更连贯)。来袭锥为结构阈值。幅度按净空缺口驱动(感知体型):只让出几何所需——对方扫掠走廊半宽(避障副轴,采集时已含旋转扫掠膨胀)+ 自身沿侧让方向厚度 + ClearanceMargin,减去自身到对方路径线的当前垂距;缺口按 ClearanceTime 折算成速度(统一 BiasSpeedFrac × 最大速度 封顶,再按质量 / 强度 / 责任加权),总量以 Arbitration.BiasCapFactor 封顶。走廊让净即停——步兵路过人群几乎不动,坦克路过让出整条通道。

Side-step maneuver primitive (replaces OncomingSideStep): IDLE self only adds a lateral DV component perpendicular to each approaching neighbor's velocity so it steps aside instead of being pushed along the lane. Returners and movers are in-transit and use LaneShift instead (in-transit agents have a real desired velocity; a lane rotation is more coherent than per-neighbor sidesteps). The approach cone is a structural threshold. Magnitude is clearance-deficit driven (size-aware): the step only covers what geometry demands — the passer's swept corridor half-width (avoidance minor axis, rotation-sweep-inflated at gather) plus self's extent along the step direction plus ClearanceMargin, minus the current lateral gap to the passer's path line; the deficit converts to speed over ClearanceTime (capped by BiasSpeedFrac × MaxSpeed, then mass/strength/yield-weighted), total capped by Arbitration.BiasCapFactor. Corridor clear → step stops — infantry passers barely nudge the crowd, tank passers get a full lane yielded.

贴墙滑行规则(ORCA 求解器前置的预解算偏置):自身 DV 有朝向障碍物分量(DV·N < 0)时,将速度向墙壁切线方向投影。强度按环形区缩放:身体接触(Dist ≤ SelfRadius)→ Strength=1(完全贴墙滑行);避障椭圆边缘(Dist ≥ Outer)→ Strength=0(不变);中间 C0 连续。不重归一化——移除向内分量自然减速。

Wall-slide rule (pre-solve bias ahead of the ORCA solver): when self's DV has a component INTO an obstacle (DV·N < 0), the velocity is projected toward the wall tangent. Strength scales over the annular zone: at body contact (Dist ≤ SelfRadius) → Strength=1 (full slide, pure tangent); at the avoidance ellipse rim (Dist ≥ Outer) → Strength=0 (no change); C0-continuous between. No renormalization — removing the inward component decelerates naturally.

压力场人群解耦(求解器后置偏置):启用后压力梯度将求解器输出引向稀疏区域,使 Agent 主动绕开密集人群而非仅在 PBD 接触后反应。消费与物理漂移路径(Collider.bPushedByPressureField)相同的 FFix64 压力网格;本规则是避障层的按原型开关(压力场本身由沉积方 Collider.bDepositToPressureField 供源)。

Pressure-field crowd decoupling via avoidance (a post-solve bias): when enabled, the pressure gradient steers the solver output toward sparser areas so agents proactively route around dense crowds instead of only reacting after PBD contact. Consumes the same FFix64 pressure grid as the physics drift path (Collider.bPushedByPressureField); this rule is the per-archetype opt-in for the AVOIDANCE layer (the pressure field itself is fed by Collider.bDepositToPressureField on the depositing agents).

交通规则的共享 fragment(FA_MassBattleBaseSharedFragment 派生)——每个原型一份,所有 Agent 共享读取;以可变共享片段分发,允许 GetMutableSharedFragment 运行时调整。8 条规则聚合于此,各自独立开关且默认全开;总闸 FAvoidance::bUseTrafficRules(默认关)是启用整个交通层的唯一开关。DeterministicHash 是确定性 CRC32 种子,模板构建时从 asset path(GetPathName)赋值,保证双端避障行为一致。

Shared fragment for the traffic rules (derived from FA_MassBattleBaseSharedFragment) — one instance per archetype, read by all agents; distributed mutable so runtime tweaks via GetMutableSharedFragment are possible. All 8 rules aggregate here with independent toggles, all defaulting ON; the master gate FAvoidance::bUseTrafficRules (default OFF) is the single switch that arms the whole layer. DeterministicHash is a deterministic CRC32 seed, set from the asset path (GetPathName) at template build so avoidance behavior is identical across lockstep peers.

FTrafficShared 在 MassBattle 管线中的位置:由 MassBattleAgentMoveProcessor / MassBattleAgentMoveProcessorFP 的避障求解段读取(预解算机动原语、ComputeAvoidanceRatio 责任仲裁、求解器范围规则),MassBattleGridUpdateProcessor 在网格注册时读取(攻击者旗标、交通等级)。规则受 FAvoidance::bUseTrafficRules 总闸控制(默认关)。

Position of FTrafficShared in the MassBattle pipeline: read by the avoidance-solver sections of MassBattleAgentMoveProcessor / MassBattleAgentMoveProcessorFP (pre-solve maneuver primitives, ComputeAvoidanceRatio yield arbitration, solver-scope rules); MassBattleGridUpdateProcessor reads it at grid registration (attacker flags, traffic ranks). All rules are gated by FAvoidance::bUseTrafficRules (default OFF).