Roadrunner Rrhd Authoring

RRHD Authoring Skill

Build RoadRunner HD Map (.rrhd) entities directly using the roadrunnerHDMap API in MATLAB. Provides the generic building blocks for constructing RRHD from any source — synthetic scenes, converted map data, or custom formats.

When to Use

• Building RRHD maps from scratch (synthetic scenes, test roads)

• Assembling RRHD entities from converted map data (Lanelet2, HERE, NDS, OpenDRIVE, custom formats)

• Need reusable geometry algorithms: alignment detection, center line synthesis, junction polygons, endpoint snapping, closed-loop splitting

• Adding lanes, boundaries, markings, junctions, signs, barriers, or parking to an HD Map

• Need verified roadrunner.hdmap.* class/property reference and construction patterns

• Debugging RRHD construction errors (wrong property names, alignment issues)

Role in Map Conversion Pipelines

This skill provides format-agnostic RRHD construction patterns that any converter skill invokes:

Skill Boundaries

This skill owns (format-agnostic RRHD construction):

Converter skills own (source-format-specific):

Decision rule:

• "How do I build this RRHD object?" → RRHD authoring skill

• "How do I extract this from my source format?" → Converter skill

When NOT to Use

• Converting Lanelet2 .osm files — use roadrunner-convert-lanelet2-to-rrhd (which invokes this skill)

• Looking up asset paths for markings/signs/barriers — use roadrunner-asset-mapping

• Importing finished .rrhd into RoadRunner — use roadrunner-import-scene

• Editing an existing RoadRunner scene interactively (this skill writes .rrhd files, not scene edits)

Key Rules

• Always write to .m files when executing code. Never put multi-line MATLAB code directly in evaluate_matlab_code. Write to a .m file, run with run_matlab_file, edit on error. Exception: if the user asks to "show the pattern" or says "do not execute", show code inline without writing files.

• Read only the specific reference needed for the task — do not read all references upfront. Use apiReference.md for property lookups, other references only when directly relevant.

• Only build what the user asked for. Do not add extra lanes, junctions, or objects unless explicitly requested.

• rrMap = roadrunnerHDMap; must come first — loads the namespace before any roadrunner.hdmap.* usage.

• Create-then-assign pattern — never pass constructor args (except RelativeAssetPath and AlignedReference).

• Empty typed arrays — use ClassName.empty not [].

• All geometry must be Nx3 — always include Z column (default 0 if flat).

• Run enforcement gate before write() — alignment, spatial, and geometry checks are mandatory.

• Track maps (closed-loop ovals/circuits): OMIT topology (no Predecessors/Successors), OMIT marking attributes (no ParametricAttribution), OMIT Z bumps. Use natural elevation only. See references/orthogonalSplit.md.

Critical API Rules

You MUST create a roadrunnerHDMap object before using any roadrunner.hdmap.* classes (lazy namespace loading):

rrMap = roadrunnerHDMap;  % REQUIRED — loads the namespace

Create-then-assign pattern — never pass constructor args (except Name=Value for two classes):

ref = roadrunner.hdmap.Reference;
ref.ID = "myID";  % assign after creation

Name=Value constructors — ONLY RelativeAssetPath and AlignedReference accept Name=Value:

rap = roadrunner.hdmap.RelativeAssetPath(AssetPath="Assets/Markings/StopLine.rrlms");
ar = roadrunner.hdmap.AlignedReference(Reference=ref, Alignment="Forward");
% WRONG: positional args error with "A name is expected"

Empty typed arrays — never use [] for typed properties:

lane.Predecessors = roadrunner.hdmap.AlignedReference.empty;  % CORRECT
lane.Predecessors = [];  % WRONG: "Value must be of type AlignedReference"

Verified Property Names (R2025a+)

See references/apiReference.md for the complete verified class/property table. Key gotchas:

Synthetic Scene Construction

Lane Width and Geometry

Standard lane width: 3.7m (US highway). Compute boundary positions by offsetting perpendicular to the lane center line direction.

Shared Boundaries Between Adjacent Lanes

Adjacent lanes MUST share boundary geometry. The right boundary of lane N is the same object as the left boundary of lane N+1. For N lanes, create N+1 boundaries.

  LB_0 (left edge)
  ───────────────────
  │    Lane_1        │  left=LB_0(Fwd), right=LB_1(Fwd)
  ───────────────────
  LB_1 (shared)
  ───────────────────
  │    Lane_2        │  left=LB_1(Fwd), right=LB_2(Fwd)
  ───────────────────
  LB_2 (right edge)

Alignment Rules

See references/alignmentRules.md for complete rules, diagrams, and green-surface debugging.

Alignment specifies how boundary geometry direction relates to lane geometry direction:

• Same direction → "Forward"

• Opposite direction → "Backward"

Which algorithm to use:

• Map conversion (left/right known from source, e.g., Lanelet2 roles) → dp-based approach below

• Synthetic scenes (left/right must be determined from geometry) → Multi-sample spatial verification in references/alignmentRules.md

Dp-based algorithm (for conversion — left/right already known):

lDir = leftPts(end,:) - leftPts(1,:);
rDir = rightPts(end,:) - rightPts(1,:);
dp = dot(lDir(1:2)/(norm(lDir(1:2))+1e-10), rDir(1:2)/(norm(rDir(1:2))+1e-10));

if dp >= -0.3
    % Normal: both boundaries go same direction
    leftAlign = "Forward"; rightAlign = "Forward";
else
    % Opposing boundaries: use proximity to determine which is backward
    d_ls_re = norm(leftPts(1,1:2) - rightPts(end,1:2));
    d_ls_rs = norm(leftPts(1,1:2) - rightPts(1,1:2));
    if d_ls_re < d_ls_rs
        leftAlign = "Forward"; rightAlign = "Backward";
    else
        leftAlign = "Backward"; rightAlign = "Forward";
    end
end

The -0.3 threshold (not 0) handles lanes with slight boundary curvature that produce small negative dot products but are NOT truly opposing.

Left/Right spatial verification (for synthetic scenes where left/right must be verified):

nSamples = min(5, size(centerGeom,1)-1);
sampleIdx = round(linspace(2, size(centerGeom,1)-1, nSamples));
leftOnLeft = 0;
for si = 1:numel(sampleIdx)
    idx = sampleIdx(si);
    localTan = centerGeom(min(idx+1,end),1:2) - centerGeom(max(idx-1,1),1:2);
    localTan = localTan / norm(localTan);
    localLeftN = [-localTan(2), localTan(1)];
    distsL = vecnorm(leftBndGeom(:,1:2) - centerGeom(idx,1:2), 2, 2);
    [~, closestL] = min(distsL);
    toBndL = leftBndGeom(closestL,1:2) - centerGeom(idx,1:2);
    if dot(toBndL, localLeftN) > 0, leftOnLeft = leftOnLeft + 1; end
end
assert(leftOnLeft >= numel(sampleIdx)/2, ...
    'Left boundary is spatially on the RIGHT — swap boundaries or fix assignment');

Wrong alignment or swapped left/right causes green grass instead of road surface.

Standard Marking Assets

Multi-Lane Road Recipe

1. Compute N lane center lines (evenly spaced at laneWidth intervals)

2. Compute N+1 boundary lines (edges + between each lane pair)

3. Assign shared boundaries — same-direction: lane i uses left=LB_(i-1), right=LB_i. Opposing-traffic lanes: swap left/right (driver's left is toward the road edge, not the center)

4. Detect alignment via dot product (boundaries going same direction as lane → Forward, opposite → Backward)

5. Apply markings (edges: SolidSingleWhite, divider: SolidDoubleYellow, separators: DashedSingleWhite)

6. Wire topology (if multiple segments, connect with pred/succ)

7. Assemble map and write

Additional steps for converted maps only (skip for synthetic scenes):

• Left-gap topology filter — see references/leftGapFilter.md

• Snap connected endpoints — see scripts/snapConnectedEndpoints.m

• Resolve height conflicts — see scripts/resolveHeightConflicts.m

• Deduplicate boundaries — see references/deduplicateBoundaries.md

Complete 2-Lane Road Example

rrMap = roadrunnerHDMap;
w = 3.7;

% Lane Markings
lm1 = roadrunner.hdmap.LaneMarking; lm1.ID = "SolidWhite";
rap1 = roadrunner.hdmap.RelativeAssetPath; rap1.AssetPath = "Assets/Markings/SolidSingleWhite.rrlms";
lm1.AssetPath = rap1;
lm2 = roadrunner.hdmap.LaneMarking; lm2.ID = "DashedWhite";
rap2 = roadrunner.hdmap.RelativeAssetPath; rap2.AssetPath = "Assets/Markings/DashedSingleWhite.rrlms";
lm2.AssetPath = rap2;

% Boundaries (3 for 2 lanes)
lb = roadrunner.hdmap.LaneBoundary;
lb(2) = roadrunner.hdmap.LaneBoundary;
lb(3) = roadrunner.hdmap.LaneBoundary;
lb(1).ID = "LB_0"; lb(1).Geometry = [0 w 0; 100 w 0];
lb(2).ID = "LB_1"; lb(2).Geometry = [0 0 0; 100 0 0];
lb(3).ID = "LB_2"; lb(3).Geometry = [0 -w 0; 100 -w 0];

% Markings on boundaries
for i = [1 3]  % solid on edges
    pa = roadrunner.hdmap.ParametricAttribution; pa.Span = [0 1];
    ref = roadrunner.hdmap.Reference; ref.ID = "SolidWhite";
    mr = roadrunner.hdmap.MarkingReference; mr.MarkingID = ref; mr.FlipLaterally = false;
    pa.MarkingReference = mr;
    lb(i).ParametricAttributes = pa;
end
pa = roadrunner.hdmap.ParametricAttribution; pa.Span = [0 1];
ref = roadrunner.hdmap.Reference; ref.ID = "DashedWhite";
mr = roadrunner.hdmap.MarkingReference; mr.MarkingID = ref; mr.FlipLaterally = false;
pa.MarkingReference = mr;
lb(2).ParametricAttributes = pa;

% Lanes
ln = roadrunner.hdmap.Lane;
ln(2) = roadrunner.hdmap.Lane;
ln(1).ID = "Lane_1"; ln(1).Geometry = [0 w/2 0; 100 w/2 0];
ln(1).TravelDirection = "Forward"; ln(1).LaneType = "Driving";
ln(2).ID = "Lane_2"; ln(2).Geometry = [0 -w/2 0; 100 -w/2 0];
ln(2).TravelDirection = "Forward"; ln(2).LaneType = "Driving";

% Assign boundaries with alignment
bndIDs = ["LB_0","LB_1","LB_2"];
for i = 1:2
    arL = roadrunner.hdmap.AlignedReference;
    refL = roadrunner.hdmap.Reference; refL.ID = bndIDs(i);
    arL.Reference = refL; arL.Alignment = "Forward";
    ln(i).LeftLaneBoundary = arL;

    arR = roadrunner.hdmap.AlignedReference;
    refR = roadrunner.hdmap.Reference; refR.ID = bndIDs(i+1);
    arR.Reference = refR; arR.Alignment = "Forward";
    ln(i).RightLaneBoundary = arR;
end

% Assemble and write
rrMap.Lanes = ln;
rrMap.LaneBoundaries = lb;
rrMap.LaneMarkings = [lm1 lm2];
write(rrMap, "two_lane_road.rrhd");

Enforcement Gate (MANDATORY before write())

Run this validation block before writing any multi-lane map. Do NOT skip. If any check FAILS, fix the issue (run the appropriate post-processing step) before proceeding to write().

%% --- ENFORCEMENT: Topology — no excessive connections (left-gap filter applied) ---
lanes = rrMap.Lanes;
nExcessiveJunc = 0;
for i = 1:numel(lanes)
    nConn = numel(lanes(i).Predecessors) + numel(lanes(i).Successors);
    if nConn > 4  % junction lanes should have at most ~4 (multi-lane merge/diverge)
        nExcessiveJunc = nExcessiveJunc + 1;
    end
end
if nExcessiveJunc > 0
    warning('TOPOLOGY WARNING: %d lanes have >4 connections — run left-gap filter (references/leftGapFilter.md)', nExcessiveJunc);
end
fprintf('Topology check: %d lanes with >4 connections\n', nExcessiveJunc);

%% --- ENFORCEMENT: Alignment computed (not all Forward) ---
if numel(lanes) > 2
    nFwd = 0; nBwd = 0;
    for i = 1:numel(lanes)
        if lanes(i).LeftLaneBoundary.Alignment == "Forward", nFwd = nFwd+1;
        else, nBwd = nBwd+1; end
        if lanes(i).RightLaneBoundary.Alignment == "Forward", nFwd = nFwd+1;
        else, nBwd = nBwd+1; end
    end
    assert(nBwd > 0 || numel(lanes) <= 2, ...
        'ALIGNMENT ERROR: All boundaries Forward for %d lanes — compute dot product per lane.', numel(lanes));
    fprintf('Alignment: %d Forward, %d Backward — OK\n', nFwd, nBwd);
end

%% --- ENFORCEMENT: Left boundary spatially on left ---
bnds = rrMap.LaneBoundaries;
bndMap = containers.Map;
for i = 1:numel(bnds), bndMap(bnds(i).ID) = bnds(i).Geometry; end
nBadSide = 0;
for i = 1:numel(lanes)
    lGeom = lanes(i).Geometry;
    lDir = lGeom(end,1:2) - lGeom(1,1:2);
    lDir = lDir / norm(lDir);
    leftN = [-lDir(2), lDir(1)];
    leftBndID = lanes(i).LeftLaneBoundary.Reference.ID;
    if bndMap.isKey(leftBndID)
        leftBndGeom = bndMap(leftBndID);
        toBnd = mean(leftBndGeom(:,1:2)) - mean(lGeom(:,1:2));
        if dot(toBnd, leftN) < 0, nBadSide = nBadSide + 1; end
    end
end
assert(nBadSide == 0, ...
    'SPATIAL ERROR: %d lanes have left boundary on wrong side — fix alignment algorithm.', nBadSide);
fprintf('Spatial left-side check: PASS\n');

%% --- ENFORCEMENT: Geometry is Nx3 ---
bnds = rrMap.LaneBoundaries;
for i = 1:numel(bnds)
    assert(size(bnds(i).Geometry, 2) == 3, ...
        'Boundary %s geometry must be Nx3 (got Nx%d)', bnds(i).ID, size(bnds(i).Geometry,2));
end
fprintf('Geometry dimensions: PASS\n');

%% --- ENFORCEMENT: GeoReference set ---
assert(any(rrMap.GeoReference ~= 0), 'GeoReference is [0,0] — set lat/lon origin');
fprintf('GeoReference: PASS\n');

%% --- ENFORCEMENT: No duplicate boundaries (deduplication applied) ---
bnds = rrMap.LaneBoundaries;
nDupBnd = 0;
for i = 1:numel(bnds)
    for j = i+1:numel(bnds)
        g1 = bnds(i).Geometry; g2 = bnds(j).Geometry;
        if size(g1,1) ~= size(g2,1), continue; end
        if max(vecnorm(g1 - flipud(g2), 2, 2)) < 0.001
            nDupBnd = nDupBnd + 1; break;
        end
    end
    if nDupBnd > 5, break; end  % early exit for performance
end
if nDupBnd > 0
    warning('DEDUP WARNING: Found reversed-duplicate boundaries — run deduplicateBoundaries (references/deduplicateBoundaries.md)');
end
fprintf('Boundary deduplication check: %d duplicates found\n', nDupBnd);

Post-Processing (MANDATORY — execute in order before write())

Left-Gap Topology Filter (MANDATORY for converted maps)

Removes false predecessor/successor connections caused by shared boundary endpoint nodes. For each lane with multiple successors (or predecessors), computes left boundary gap — if any connection has gap < 3m and others > 3m, removes the far ones. Without this, junction lanes get 2-4x too many connections. See references/leftGapFilter.md.

Endpoint Snapping

For connected lanes (pred/succ), snap successor start to predecessor end. See scripts/snapConnectedEndpoints.m and references/snapEndpoints.md.

Height Conflict Resolution

Overlapping unconnected lanes at the same Z cause grass artifacts. Use graph coloring to assign Z-levels. See scripts/resolveHeightConflicts.m and references/resolveHeights.md.

Boundary Deduplication (MANDATORY for converted maps)

When building from source formats that assign separate boundary objects to opposing lanes (Lanelet2, HERE, NDS, etc.), scan all boundary pairs for identical geometry (forward or reversed within 1mm tolerance). Merge duplicates and update lane references with flipped alignment. Without this step, shared boundaries render doubled markings. See references/deduplicateBoundaries.md.

Additional Entity Types

For all entity classes (Signs, Signals, Barriers, Parking, CurveMarkings, StencilMarkings, StaticObjects, LaneGroups), see the class/property table in references/apiReference.md. Key references by task:

Key Functions

Known Limitations

• SignalTypes/Signals survive RRHD write/read in MATLAB, but RoadRunner silently ignores them on scene import — signals will not appear in the RoadRunner scene

• RoadRunner cannot render self-closing lanes — split using N-way orthogonal algorithm (see references/orthogonalSplit.md)

• Geometry must always be Nx3 (include Z=0 if flat)

• For track maps (closed-loop ovals): omit topology (Predecessors/Successors), Z bumps, and marking attributes (ParametricAttribution) — see orthogonalSplit.md Track-Specific Guidance

Conventions

• Always use create-then-assign pattern for roadrunner.hdmap.* objects

• Use RelativeAssetPath(AssetPath="...") and AlignedReference(Reference=ref, Alignment="...") — Name=Value only

• Use ClassName.empty for empty typed arrays, never []

• All geometry is Nx3 with Z column (default 0 for flat terrain)

• Shared boundaries: one object per way, multiple lanes reference with different alignments

Copyright 2026 The MathWorks, Inc.