Configuration Details¶

All settings are stored in UserDefaults and can be exported, imported, or reset via the Data tab — see Data Tab — Maps & Config.

ROS2 bridge (Zenoh)¶

Key	Type	Default	Description
`zenohIsRouter`	Bool	`false`	`false` = Client (connects out to a router). `true` = Router (phone listens on :7447).
`zenohLocator`	String	—	Client mode only: address of the Zenoh router, e.g. `tcp/192.168.1.100:7447`.
`autoConnect`	Bool	`true`	Auto-connect to Zenoh on app launch.

Power impact of Zenoh mode:

Mode	Power draw	Notes
Router (`zenohIsRouter = true`)	Higher	Phone runs a Rust TCP listener on :7447 and manages all peer routing. Background threads run continuously even when no robot is connected.
Client (`zenohIsRouter = false`)	Lower	Phone dials out to an external router and only maintains one outbound session. Preferred for battery life when a router (e.g. Raspberry Pi) is already on the network.

Foxglove WebSocket server¶

The Foxglove WebSocket server (foxgloveEnabled) streams all topics to Foxglove Studio over a local WebSocket connection on port 8765. It runs independently of Zenoh — you can use it with or without a Zenoh connection.

Key	Type	Default	Description
`foxgloveEnabled`	Bool	`false`	Enable the Foxglove WebSocket server. Advertises all topics to Foxglove Studio on the local network.

Power impact: The Foxglove server has a measurable power cost even when no client is connected, because the listener socket is held open and every published message is serialised (CDR-encoded) and dispatched to all subscribers. With an active Foxglove Studio connection receiving high-rate topics (camera, point cloud, IMU), the additional CPU load can be significant — comparable to adding another Zenoh peer. Recommendations:

Disable when not actively debugging in Foxglove.
On battery, prefer enabling Foxglove briefly to capture a session rather than leaving it on continuously.
High-bandwidth topics (/camera/image_raw/compressed, /camera/depth/points, /cloud_map) are the largest contributors — unsubscribe from panels you are not viewing.

Operating mode¶

Key	Type	Default	Values
`appMode`	String	`"sensorBridge"`	`"sensorBridge"` · `"autonomousIDD"`

See Operating Modes for what each mode does.

SLAM quality (RTABMap)¶

Tuning difficulty: RTAB-Map has hundreds of interdependent parameters. The ones exposed here cover the highest-impact knobs for mobile LiDAR SLAM, but finding the right combination for a specific environment takes iteration. Start with defaults and change one parameter at a time, then Reset the map to compare. The official RTAB-Map parameter guide and the RTAB-Map ROS wiki are the best reference for understanding how upstream defaults differ from the values chosen here.

These map directly to RTAB-Map C++ parameters fed through iROSBridgeSetCloudParams and iROSBridgeSetSLAMParams. Changes apply immediately without a restart; the cloud rebuild happens on the next SLAM node (every 5 nodes visible in /cloud_map).

Input filtering¶

Key	Type	Default	RTAB-Map param	Description
`rtabmapEnabled`	Bool	`true`	—	Enable RTAB-Map loop closure. Off = ARKit VIO only: no loop closure, no `/cloud_map`, no map saving. Lighter on CPU (~30% less). Use when environment is dark or featureless.
`rtabMaxDepth`	Double	`2.0` m	`RGBD/DepthMax`	Clips depth measurements beyond this range before feeding SLAM. Range 1.0–5.0 m.
`rtabVoxelSize`	Double	`0.05` m	`Grid/CellSize` / voxel filter	Downsamples the point cloud to one point per voxel cell before SLAM processing. `0.0` = disabled (no downsampling).
`rtabDecimation`	Int	`4`	depth decimation	Keeps 1 in N depth points fed to SLAM. Higher = fewer points, less CPU, coarser cloud.
`rtabMinConf2`	Int	`1`	ARKit confidence mask	Filters ARKit LiDAR pixels by confidence before SLAM: `0` = all pixels, `1` = medium + high only, `2` = high confidence only.

Map quality tradeoffs:

Goal	Recommended changes
Better wall detail / denser cloud	`rtabDecimation` → 2, `rtabVoxelSize` → 0.02–0.03 m
Cleaner cloud, fewer floaters	`rtabMinConf2` → 1 or 2; enable noise filter (see below)
Larger room coverage	`rtabMaxDepth` → 3.0–4.0 m
Reduce CPU / thermal throttle	`rtabDecimation` → 6–8, `rtabVoxelSize` → 0.05 m, `rtabMaxDepth` → 1.5 m

Node insertion rate¶

Key	Type	Default	RTAB-Map param	Description
`rtabLinearUpdate`	Double	`0.05` m	`RGBD/LinearUpdate`	Minimum travel (metres) before RTAB-Map adds a new SLAM node. Upstream default is 0.1 m; app uses 0.05 m for denser nodes.
`rtabAngularUpdate`	Double	`0.05` rad	`RGBD/AngularUpdate`	Minimum angular change (radians) before a new SLAM node is added. Upstream default is 0.1 rad.

Lower values → more nodes → more accurate loop closure but more memory and CPU. For slow-moving robots keep at 0.05 m / 0.05 rad. For fast sweeps raise to 0.10 m / 0.10 rad.

Cloud map¶

Key	Type	Default	Description
`cloudMapEnabled`	Bool	`false`	Publish the full accumulated `/cloud_map` on every SLAM event. Expensive at scale (serialises the entire cloud over the bridge each time). Use `/camera/depth/points` for incremental updates instead. Gated at the C++ level — no CPU cost when disabled.

SLAM debug (loop closure)¶

These control how RTAB-Map detects revisited places. Accessible under Settings → SLAM Debug (Loop Closure).

Live diagnostics (read-only)¶

The SLAM Debug section shows live stats updated at 1 Hz:

Field	Meaning
Loop closures	Total accepted loop closures this session. A healthy map will accumulate these as you revisit areas.
Last LC node	Node ID of the most recent loop closure. Useful for correlating with map artifacts.
LC hypothesis	Confidence of the last loop closure (0–1). Values > 0.5 are candidates; accepted closures pass additional geometric verification.
Vis inliers / matches	Inlier feature count / total candidate matches for the last loop closure attempt. Inliers / Matches ratio shows match quality — below 0.3 usually means a false positive was rejected.

Feature extraction¶

Key	Type	Default	RTAB-Map param	Description
`rtabVisMinInliers`	Int	`50`	`Vis/MinInliers`	Minimum geometric inliers required to accept a loop closure. RTAB-Map upstream default is 20; app default is 50 (stricter). Lower = more closures (risk of ghost corrections). Higher = fewer false positives but may miss real revisits. Range 5–100.
`rtabVisMaxFeatures`	Int	`400`	`Vis/MaxFeatures`	Maximum keypoints extracted per frame for loop closure detection. Upstream default is 1000; app default is 400 to save CPU. Higher → better recall in repetitive environments (long corridors) at the cost of ~2× CPU for feature extraction. Options: 200 / 400 / 600 / 800.
`rtabVisFeatureType`	Int	`6`	`Vis/FeatureType`	Feature detector + descriptor used for visual place recognition.

Feature type options and tradeoffs:

Value	Name	CPU	Quality	Best for
`6`	GFTT/BRIEF (default)	Lowest	Good	Room-scale, well-lit, fast operation
`2`	ORB	Low	Better under blur	Motion blur, faster movement
`7`	BRISK	Medium	Rotation-invariant	Environments with large viewpoint changes
`9`	ORB-Octree	Medium	More uniform coverage	Large uniform walls with few texture features

Note: Changing feature type requires resetting the map (Scan tab → Reset). The visual dictionary is incompatible between detector types.

Image decimation¶

Key	Type	Default	RTAB-Map param	Description
`rtabImagePreDecim`	Int	`2`	`Mem/ImagePreDecimation`	Downscales the RGB image before feature extraction. `1` = full resolution, `2` = half (default), `4` = quarter. Upstream default is 1 (no decimation); app default is 2 to halve feature extraction CPU. Full res (`1`) improves loop closure reliability in large or repetitive spaces.

Frame rate to SLAM¶

Key	Type	Default	Description
`rtabSkip`	Int	`10`	Feed 1-in-N ARKit frames (~30 Hz) to RTAB-Map. `10` ≈ 3 Hz to SLAM. Lower = more frames → better tracking and more nodes, but heavier CPU. Raise if you see "retaining ARFrames" warning in logs (means SLAM is slower than ARKit delivery).

CPU impact of rtabSkip:

Value	SLAM Hz	CPU cost	Use when
`1`	30 Hz	Very high	Debugging only
`5`	6 Hz	High	Slow robot, need dense nodes
`10`	3 Hz	Medium (default)	Normal operation
`20`	1.5 Hz	Low	Large rooms, fast movement
`60`	0.5 Hz	Very low	Thermal relief only

Point cloud noise filter¶

Applied after depth unprojection, before occupancy grid and SLAM input.

Key	Type	Default	Description
`cloudNoiseRadius`	Double	`0.0` m	Radius for neighbor search. `0` = disabled. Good starting value: `0.05` m. Each point with fewer than `cloudNoiseMinNeighbors` neighbors within this radius is removed.
`cloudNoiseMinNeighbors`	Int	`1`	Points with fewer neighbors than this within the search radius are discarded. Raise to 2–4 for aggressive filtering.
`gridRayTracing`	Bool	`true`	Enables free-space ray tracing: clears occupancy voxels between the sensor origin and each depth measurement. Removes phantom obstacles behind surfaces and stale occupied cells as the robot moves. Slight CPU cost per frame; keep enabled for navigation.
`gridNormalsSegmentation`	Bool	`true`	Classifies depth points as ground/obstacle using surface normals (angle relative to gravity) instead of height thresholds alone. Detects stair risers and drop-offs. Can misfire with sparse/decimated depth — if the occupancy grid shows noise on flat floors, switch to height-only (disable).

Tuning the noise filter:

Floaters above the floor → enable filter, radius = 0.05 m, minNeighbors = 2. Walls look eroded → filter is too aggressive; lower minNeighbors to 1 or increase radius. Floor classified as obstacle → disable gridNormalsSegmentation or tune camera height parameters.

Tips: getting a better map¶

More loop closures¶

Move slowly and steadily — RGBD/LinearUpdate adds nodes based on travel; fast sweeps produce sparse graphs.
Revisit areas from similar angles — loop closure is view-dependent. Approaching from the opposite direction helps.
Raise rtabVisMaxFeatures to 600–800 in repetitive or textureless environments (long corridors, white walls).
Lower rtabVisMinInliers to 30–40 if the live stats show high match counts but LC hypothesis never crosses 0.5.

Cleaner point cloud¶

Enable the noise filter with cloudNoiseRadius = 0.05 m, cloudNoiseMinNeighbors = 2.
Keep gridRayTracing enabled — it removes ghost points left behind as the robot moves.
Use rtabMinConf2 = 1 (medium+) to skip low-confidence LiDAR pixels; step up to 2 (high only) in cluttered spaces.
Lower rtabVoxelSize to 0.02–0.03 m for denser output (trades CPU).

Reduce CPU / prevent thermal throttle¶

Raise rtabSkip to 15–20 (drops SLAM Hz, not ARKit tracking Hz).
Raise rtabDecimation to 6–8.
Raise rtabVoxelSize to 0.05–0.10 m.
Set rtabImagePreDecim to 4 (quarter resolution for feature extraction).
Disable cloudMapEnabled — rebuilding the full cloud every node is expensive at scale.

Tracking CPU usage¶

iROSLink publishes live diagnostics to /diag/cpu_percent and /diag/ram_mb (std_msgs/Float32) and mirrors them to Foxglove over WebSocket.

In Foxglove Studio: - Add a Plot panel → subscribe to /diag/cpu_percent and /diag/ram_mb. - Watch for CPU spikes > 80% sustained — this precedes thermal throttle and ARKit frame drops.

In-app diagnostics (Topics tab):

Topic	Type	What it shows
`/diag/cpu_percent`	Float32	Process CPU % (sampled every 10 s)
`/diag/ram_mb`	Float32	App RAM usage in MB
`/diag/speed_cmps`	Float32	Forward speed cm/s (from ARKit pose diff)
`/diag/yaw_rate_dps`	Float32	Yaw rate deg/s

Signs of CPU pressure and remedies:

Symptom	Cause	Fix
"Retaining ARFrames" in logs	SLAM can't keep up with ARKit	Raise `rtabSkip`
Map stops updating	SLAM thread stalled	Raise `rtabDecimation`, `rtabVoxelSize`
ARKit tracking degrades	Thermal throttle	Lower `rtabMaxDepth`, disable `cloudMapEnabled`
High RAM (> 400 MB)	Growing node graph	Raise `rtabLinearUpdate` to space out nodes

Control tab keys (Autonomous IDD only)¶

→ Control Tab — Settings & Calibration — occupancy grid, robot body, camera rates, navigation, motor calibration, autonomous behaviours, gamepad buttons

App behaviour¶

Key	Type	Default	Description
`keepScreenOn`	Bool	`true`	Disable auto-lock while app is running.
`stopSessionOnComplete`	Bool	—	Auto-pause SLAM when explore or clean finishes.