Lanelet2 to RRHD Converter

Converts Lanelet2 .osm files to RoadRunner HD Map .rrhd format.

When to Use

• Converting a Lanelet2 .osm file to RoadRunner HD Map .rrhd format
• Importing Lanelet2 maps into RoadRunner via MATLAB
• Building RRHD from .osm sources that contain type=lanelet relations
• Need full pipeline: parse, geometry, topology, semantics, junctions, barriers, signs, markings

When NOT to Use

• Input is a standard OpenStreetMap file (highway=* ways without type=lanelet relations)
• Input is OpenDRIVE .xodr — import directly via roadrunner-import-scene
• Building RRHD from scratch without a source file — use roadrunner-rrhd-authoring
• Only need asset path lookups — use roadrunner-asset-mapping

Key Rules

• Always write to .m files. Never put multi-line MATLAB code directly in evaluate_matlab_code. Write to a .m file, run with run_matlab_file, edit on error.
• ALL pipeline steps are mandatory. Do NOT stop after writing lanes/boundaries — junctions, curve markings, barriers, signs, and speed limits must all be built.
• Boundary geometry is IMMUTABLE. Never flip, resample, project, or modify boundary points.
• One boundary object per way. Deduplicate via wayToBndID map.
• Detect alignment for EVERY lane. Never hardcode "Forward" for all boundaries.
• Center line density: 1 point per meter minimum. Use max(10, round(avgLen), nLeft, nRight).
• Run enforcement gate before `write()`. Alignment, spatial, extension, and completeness checks are mandatory.
• No nested function definitions. Code runs in script context — all logic inline or anonymous functions.

Behavior

Generate MATLAB code that performs the conversion pipeline described below. Run it via mcp__matlab__evaluate_matlab_code. No pre-built scripts or addpath calls are needed — Claude generates all code at runtime from these instructions.

MANDATORY: ALL steps must be executed. Do NOT stop after writing lanes/boundaries. The pipeline is incomplete without: junctions, curve markings (stop lines + crosswalks), barriers, signs, and speed limits. Every element found in the OSM MUST appear in the output RRHD.

MANDATORY: Steps 3b and 3c (discovery) must execute IN THE SAME code block as Step 3a. These are not optional post-processing — they are part of parsing. The discovery loop code is inlined below in Step 3. If you skip 3b/3c, the completeness gate in Step 9 will fail with assertion errors.

Scope: This skill supports Lanelet2 OSM only. If the input is a standard OpenStreetMap file (highway=* ways without type=lanelet relations), do NOT attempt conversion. Instead inform the user: "This file appears to be a standard OpenStreetMap road network, not a Lanelet2 map. This skill only supports Lanelet2 OSM → RRHD conversion."

Companion skills (invoke automatically during conversion):

These skills are loaded on demand — read their references when generating RRHD construction code or resolving asset paths. This skill provides the Lanelet2-specific parsing and bridging logic.

Coordinate System

Lanelet2 OSM nodes may use either:

• `local_x`/`local_y` tags — already in local meters, use directly as geometry X/Y
• `lat`/`lon` attributes only — geographic coordinates that MUST be projected to local meters

Step 1: Check for map_projector_info.yaml

Lanelet2 datasets typically include a map_projector_info.yaml file in the same directory as the .osm file. This file specifies the projection and map origin:

ProjectorType: TransverseMercator
VerticalDatum: WGS84
MapOrigin:
  Latitude: 42.300945
  Longitude: -83.698205
  Altitude: 0

Always look for this file first. Parse it to get ProjectorType and MapOrigin:

projFile = fullfile(fileparts(osmFile), "map_projector_info.yaml");
if isfile(projFile)
    txt = fileread(projFile);
    latMatch = regexp(txt, 'Latitude:\s*([-\d.]+)', 'tokens');
    lonMatch = regexp(txt, 'Longitude:\s*([-\d.]+)', 'tokens');
    if ~isempty(latMatch) && ~isempty(lonMatch)
        originLat = str2double(latMatch{1}{1});
        originLon = str2double(lonMatch{1}{1});
    end
end

If MapOrigin is [0, 0], the origin is implicit — use the centroid of all nodes instead.

Step 2: Determine coordinate mode

1. Nodes have `local_x`/`local_y` → use directly, set geoRef from map_projector_info.yaml MapOrigin (or from node lat/lon if yaml missing)
2. Nodes have only `lat`/`lon` → project to local ENU meters using the origin from yaml (or node centroid as fallback)

Step 3: Project lat/lon to local meters (when needed)

% Origin from map_projector_info.yaml or centroid fallback
geoRef = [originLat, originLon];

% For each node, convert to local meters:
dLat = node.lat - originLat;
dLon = node.lon - originLon;
metersPerDegLat = 111132.92;
metersPerDegLon = 111132.92 * cosd(originLat);
x = dLon * metersPerDegLon;   % East
y = dLat * metersPerDegLat;   % North
z = node.ele;                  % Up (elevation)

Set rrMap.GeoReference = [originLat, originLon] so RoadRunner knows the map's geographic position.

IMPORTANT: This projection happens ONCE during node parsing. All downstream geometry (boundaries, center lines, barriers, signs) uses the projected local coordinates. The "boundary geometry is immutable" rule (below) refers to the projected coordinates — do not re-project or modify them after initial parsing.

Geometry Invariants (MUST enforce — violations produce broken RRHD)

These rules are NON-NEGOTIABLE. Every conversion MUST follow them:

1. Boundary geometry is IMMUTABLE. Store way node coordinates (local or projected) exactly as computed during parsing. NEVER flip, resample, or modify boundary points after initial coordinate assignment. Shared boundaries between opposing lanes will corrupt if touched. Use Alignment="Backward" instead.

1. One boundary object per way. Deduplicate via wayToBndID map. Multiple lanes reference the same boundary with different alignments.

1. Alignment via proximity + multi-sample spatial verification. Use the robust algorithm in roadrunner-rrhd-authoring/references/alignmentRules.md: (1) compare boundary start/end distance to lane start for direction, (2) verify with local tangent at multiple sample points. NEVER rely on a single overall direction dot product — it fails on curved segments.

1. Center line density: 1 point per meter minimum. Use nPts = max(10, round(avgBoundaryLength), nLeftPts, nRightPts). Too few points causes pchip interpolation deviation from true midpoint.

1. Center line orthogonal endpoints only. Enforce perpendicular start/end on CENTER LINE via tangent blending. Do NOT project boundary endpoints.

MATLAB Script Constraints

Code runs in mcp__matlab__evaluate_matlab_code which is a script context (NOT a function file):

• NO nested function definitions — all logic must be inline or use anonymous functions
• NO `function ... end` blocks — these error with "Function definitions are not supported in this context"
• Empty typed arrays — use ClassName.empty not [] (e.g., roadrunner.hdmap.AlignedReference.empty)
• Name=Value constructors — RelativeAssetPath, AlignedReference require Name=Value syntax:

  % CORRECT:
  rap = roadrunner.hdmap.RelativeAssetPath(AssetPath="Assets/Markings/StopLine.rrlms");
  ar = roadrunner.hdmap.AlignedReference(Reference=ref, Alignment="Forward");

  % WRONG (will error with "A name is expected"):
  rap = roadrunner.hdmap.RelativeAssetPath("Assets/Markings/StopLine.rrlms");
  ar = roadrunner.hdmap.AlignedReference(bndID, "Forward");

• containers.Map empty assignment — never do map(key) = []. Use map(key) = zeros(1,0) or only assign non-empty values
• No chaining `()()` — geomMap(key)(:,1) errors with "Using parentheses directly after parentheses is disallowed". Assign to a temp variable first:

  % WRONG: min(geomMap(k)(:,1))
  % CORRECT:
  pts = geomMap(k);
  minX = min(pts(:,1));

• `unique()` on cell arrays — unique(cellArray) requires cellstr (cell of char vectors). For string arrays use unique(stringArray) directly. For mixed cells: unique(string(cellArray)).
• `containers.Map` values need explicit type cast — values retrieved from containers.Map may lose their numeric type. Always cast before math: double(geoRef(1)) before passing to cosd(), sind(), etc.
• Cell array struct field assignment — never do cellArray{end+1} = x; cellArray{end}.field = val; (errors with "dot indexing not supported for double"). Instead, build the struct first then append:

  % WRONG:
  stopLineWays{end+1} = w; stopLineWays{end}.wayID = wid;
  % CORRECT:
  w.wayID = wid; stopLineWays{end+1} = w;
  % OR: store IDs in a separate parallel cell array

Pipeline (execute in order)

Step 1: Parse OSM

See references/readOSM.md and scripts/parseOSM.m.

doc = xmlread(osmFile);
% Parse nodes: id, lat, lon, ele, local_x, local_y (from tags)
% Parse ways: id, nodeRefs[], tags (containers.Map)
% Parse relations: id, members[], tags (containers.Map)

Step 2: Validate Format

Check that at least one relation has type=lanelet. If none found, error with clear message identifying the file as non-Lanelet2.

hasLanelet = false;
relKeys = relations.keys;
for i = 1:numel(relKeys)
    rel = relations(relKeys{i});
    if rel.tags.isKey('type') && strcmp(rel.tags('type'), 'lanelet')
        hasLanelet = true; break;
    end
end
if ~hasLanelet
    error('UNSUPPORTED FORMAT: This converter supports Lanelet2 OSM files only.');
end

Step 3: Extract Lanelets + Discover ALL Non-Lane Elements

This step has THREE mandatory sub-steps. ALL must execute.

#### Step 3a: Extract Lanelets

Filter relations with type=lanelet. See references/extractLanelets.md.

% For each relation with type=lanelet:
%   leftWayID  = member with role="left"
%   rightWayID = member with role="right"
%   subtype, one_way, speed_limit, turn_direction from tags

#### Step 3b: Discover ALL Way Types (MANDATORY — run immediately after 3a)

See scripts/discoverWayTypes.m for the reference implementation.

DO NOT use a whitelist. Scan every way, categorize by type tag using strsplit(wType, '/') for compound types:

stopLineWays = {}; pedestrianMarkingWays = {}; zebraMarkingWays = {};
bikeMarkingWays = {}; zigzagWays = {};
fenceWays = {}; guardRailWays = {}; jerseyBarrierWays = {};
wallWays = {}; curbstoneWays = {};
trafficSignWays = {}; trafficLightWays = {};
unmappedWays = containers.Map;

wayKeys = ways.keys;
for i = 1:numel(wayKeys)
    w = ways(wayKeys{i});
    if ~w.tags.isKey('type'), continue; end
    wType = w.tags('type');
    typeParts = strsplit(wType, '/');
    baseType = typeParts{1};
    switch baseType
        case 'stop_line', stopLineWays{end+1} = w;
        case 'pedestrian_marking', pedestrianMarkingWays{end+1} = w;
        case 'zebra_marking', zebraMarkingWays{end+1} = w;
        case 'bike_marking', bikeMarkingWays{end+1} = w;
        case {'zig-zag','zig_zag'}, zigzagWays{end+1} = w;
        case 'fence', fenceWays{end+1} = w;
        case 'guard_rail', guardRailWays{end+1} = w;
        case 'jersey_barrier', jerseyBarrierWays{end+1} = w;
        case 'wall', wallWays{end+1} = w;
        case 'curbstone', curbstoneWays{end+1} = w;
        case 'traffic_sign', trafficSignWays{end+1} = w;
        case 'traffic_light', trafficLightWays{end+1} = w;
        case {'line_thin','line_thick','virtual','road_border','rail','keepout','symbol'}
            % Lane boundary types — handled via lanelet extraction
        otherwise
            if unmappedWays.isKey(wType), unmappedWays(wType) = unmappedWays(wType)+1;
            else, unmappedWays(wType) = 1; end
    end
end

#### Step 3c: Discover ALL Relation Types (MANDATORY — run immediately after 3b)

See references/discoverRelations.md for the full categorization loop. Sorts all non-lanelet relations into: trafficSignRels, speedLimitRels, rightOfWayRels, trafficLightRels, multipolygonRels (struct with .building, .parking, .vegetation, .traffic_island, .walkway, .exit, .keepout), and unmappedRels (containers.Map with counts).

Print discovery summary and unmapped elements. Never silently drop anything.

Step 4: Build Geometry

Resolve way geometry from nodes. See references/extractGeometry.md for mandatory implementation patterns.

GeoReference auto-detection: If no local_x/local_y tags exist, compute centroid of all node lat/lon. Never use [0 0]. See references/detectGeoReference.md.

Priority: Use local_x/local_y node tags if available. Fall back to lat/lon → ENU (see references/latlon2enu.md):

east  = (lon - geoRef(2)) * 111320 * cosd(geoRef(1));
north = (lat - geoRef(1)) * 110540;

MANDATORY: Handle closed-loop boundaries BEFORE direction detection. If both boundaries are closed loops (start≈end gap < 1.0m), the overall direction vector is near-zero and dot(leftOverall, rightOverall) is unreliable (~0). This causes incorrect flips that produce center lines cutting across the track interior (e.g., 136m "width" instead of 6.5m). See roadrunner-rrhd-authoring skill's references/splitClosedGeometry.md for the full closed-loop boundary handling algorithm (local direction detection + start-point alignment).

MANDATORY center line synthesis pipeline — see roadrunner-rrhd-authoring skill's references/synthesizeCenterLine.md for the full algorithm with code. Summary:

1. Detect opposing boundaries: Use closed-loop-aware detection (see splitClosedGeometry.md above)
2. Resample to density-based point count using arc-length pchip — max(10, round(avgLen), nLeft, nRight)
3. Average resampled boundaries for center line
4. Enforce orthogonal endpoints with 50% tangent blending (skip for closed-loop lanes)
5. Do NOT modify boundary geometry — use Alignment property instead

CRITICAL: Do NOT apply orthogonal endpoint projection to boundary geometry. Boundaries are shared between multiple lanes (including opposing-direction lanes). Modifying a shared boundary's endpoints to match one lane's center line direction will corrupt it for other lanes referencing the same boundary. The Alignment property handles the direction relationship.

Boundary geometry rule: Always store the ORIGINAL way geometry from OSM nodes. Compute alignment using the robust multi-sample algorithm (see roadrunner-rrhd-authoring/references/alignmentRules.md). Do NOT use a simple dot product of overall directions — this fails on curved geometry.

All geometry must be Nx3 (include Z from ele tag, default 0).

Step 5: Split Closed Roads

RoadRunner cannot render self-closed geometry. See references/splitClosedRoads.md and roadrunner-rrhd-authoring skill's references/splitClosedGeometry.md for the generic N-way orthogonal split algorithm.

Split using perpendicular cross-section projection at ALL joints (including closure). Use roadrunner-rrhd-authoring/references/orthogonalSplit.md for the full N-way algorithm.

Step 6: Extract Topology

Use boundary endpoint node matching as primary method, with geometry proximity (1.0m threshold) as fallback. See references/extractTopology.md.

Critical: Account for opposing boundary directions when matching nodes.

Step 7: Detect Junctions

See references/extractJunctions.md for the full algorithm with code.

Primary method: Use turn_direction tag (straight, left, right) as the junction lane indicator. Fallback: Lane is junction if (predecessor has >1 successors) OR (successor has >1 predecessors).

Critical rules:

• BFS clustering with 10m threshold — NOT union-find (causes transitive chaining across distant intersections)
• Trace outer boundary polylines for polygon — NOT boundary(), convhull, or spline smoothing
• Min 3 lanes per cluster to avoid false positives
• Strip markings from all junction lane boundaries (don't assign ParametricAttributes)
• Graceful fallback: If detection is unreliable, skip junctions entirely — use UseLaneGroups=true in build options during import

Step 8: Extract Semantics

See references/extractSemantics.md.

Travel direction:

• turn_direction tag present → "Forward" (always unidirectional)
• one_way=yes → "Forward"
• Sidewalk/Curb → "Undirected"
• Otherwise → "Bidirectional"

Step 8b: Filter Crosswalks/Walkways from Lane List

BEFORE building RRHD lane objects, separate lanelets by subtype:

• subtype=crosswalk or subtype=walkway → route to CurveMarkings (use center line as geometry, type = SimpleCrosswalk)
• All other subtypes → build as RRHD Lanes

Never create a Lane object for crosswalk/walkway lanelets. They become CurveMarking instances.

CurveMarkingType extensions — do NOT confuse:

• Stop lines → .rrlms (lane marking style): Assets/Markings/StopLine.rrlms
• Crosswalks → .rrcws (crosswalk style): Assets/Markings/SimpleCrosswalk.rrcws

Using .rrcws for stop lines causes "Asset file is missing" on import.

Step 9: Build RRHD Objects

See references/buildRRHD.md for complete construction patterns.

Critical rules:

• rrMap = roadrunnerHDMap; must come first (loads namespace). WARNING: On some systems, roadrunnerHDMap may launch a background RoadRunner instance. The import skill's connection logic handles this with retry — see roadrunner-import-scene. Do NOT call roadrunner(InstallationFolder=...) separately if roadrunnerHDMap already launched one.
• Create-then-assign pattern for all roadrunner.hdmap.* objects
• Strip markings from junction lane boundaries (simply don't assign ParametricAttributes)
• NEVER modify boundary geometry — store original OSM way node coordinates exactly as-is. Boundaries are shared between multiple lanes (including opposing-direction lanes). Any modification (flipping, endpoint projection, resampling) will corrupt the boundary for other lanes.
• One boundary per way — use wayToBndID map to deduplicate. First lanelet to reference a way creates the boundary; all others share it via alignment.
• Detect boundary alignment FOR EVERY LANE using the robust multi-sample algorithm in roadrunner-rrhd-authoring skill's references/alignmentRules.md. NEVER hardcode `alignment = "Forward"` for all boundaries — this is the #1 cause of bad scenes. NEVER use overall direction (`geom(end,:)-geom(1,:)`) for spatial verification — it fails on highly curved segments (e.g., split half-ovals where the chord direction doesn't represent travel direction). The algorithm has two steps: (1) proximity-based direction detection, (2) multi-sample local-tangent spatial verification with majority vote.

Verified property names: See roadrunner-rrhd-authoring skill's references/apiReference.md for the complete class/property table. Key non-obvious mappings: CurveMarking.MarkingTypeReference (not CurveMarkingTypeID), Barrier.BarrierTypeReference (not BarrierTypeID), Sign.Geometry (GeoOrientedBoundingBox, not BoundingBox), SpeedLimit.Value + .VelocityUnit="Kph", ParametricAttribution.Span=[0 1] (not StartFraction/EndFraction), MarkingReference.MarkingID takes a Reference object (not numeric).

Critical: Filter self-referencing topology (lane predecessor/successor pointing to itself) before writing — RoadRunner will error with "connectedLane.Object != this" on import.

Step 10: Post-Processing (MANDATORY — do NOT skip)

Snap connected endpoints — successor start must match predecessor end (lane + boundaries). For each pred→succ pair with gap < 1.0m, snap succ start to pred end. See references/snapEndpoints.md and roadrunner-rrhd-authoring skill's scripts/snapConnectedEndpoints.m.

Height conflict resolution (MANDATORY): Overlapping unconnected lanes at same Z cause grass artifacts. Bump narrower lane by +0.05m per Z-level using graph coloring. See roadrunner-rrhd-authoring skill's scripts/resolveHeightConflicts.m.

Step 11: Assemble & Write

Assign all arrays to rrMap (Lanes, LaneBoundaries, LaneMarkings, SpeedLimits, Junctions, CurveMarkingTypes, CurveMarkings, BarrierTypes, Barriers, SignTypes, Signs) and set rrMap.GeoReference = geoRef (1x2 [lat, lon]). Call write(rrMap, outputFile). See references/buildRRHD.md.

Junctions are optional: If junction detection was skipped, omit rrMap.Junctions — defaults to empty. Use UseLaneGroups=true in build options during import.

Step 9 Sub-Steps — Build ALL Non-Lane RRHD Objects

Step 9 MUST build ALL non-lane elements from Steps 3b/3c. See references/extractNonLaneElements.md for code patterns. Sub-steps: 9a CurveMarkings (stop lines, crosswalks, bike markings), 9b Barriers (fence, guard rail, jersey barrier, wall, curbstone), 9c Signs (from ways AND relations — see references/signCodeMapping.md and references/mapLanelet2SignCode.md for sign type resolution), 9d SpeedLimits, 9e Placed Objects (see references/placedObjects.md for custom relation parsing).

Enforcement Gate (MANDATORY — run before write())

You MUST execute the validation block in references/enforcementGate.md. It catches the three most common conversion errors (alignment, spatial correctness, asset extensions, completeness). Do NOT skip or simplify it.

After Write

Print unmapped element counts (buildings, vegetation, signals — not importable). Then use roadrunner-import-scene skill to import. See references/importIntoRoadRunner.md for build options.

Key Functions

Conventions

• Store original OSM way node coordinates exactly as parsed — never modify boundary geometry
• Use containers.Map for all indexed lookups (nodes, ways, relations, geometry)
• All geometry is Nx3 with Z from ele tag (default 0)
• Use proximity-based BFS clustering for junctions (10m threshold), never union-find
• Trace outer boundary polylines for junction polygons — never boundary() or convhull
• Print discovery summary after parsing — never silently drop unmapped elements
• Use tiledlayout/nexttile for multi-panel figures (not subplot)

----

Copyright 2026 The MathWorks, Inc.