Neural network pipelines

Limelight supports two neural network pipeline types:

Classifier — classifies the entire camera frame into a single category (e.g., “cargo present” vs. “no cargo”).
Detector — detects and locates one or more objects within the frame, returning bounding boxes and class labels.

Classifier pipeline
Detector pipeline

Classifier pipeline

The classifier pipeline analyzes the full frame and outputs the most likely class. Use it when you need a simple yes/no or category answer about what the camera is seeing.

Quick reads (NetworkTables)

// Class name as a string (e.g., "note", "empty")
String className = LimelightHelpers.getClassifierClass("limelight");

// Class index (integer ID from the T2D array)
int classIndex = LimelightHelpers.getClassifierClassIndex("limelight");

getClassifierClass() reads from the tcclass NetworkTables key. getClassifierClassIndex() reads index 11 from the T2D array.

Full results via JSON

LimelightHelpers.LimelightResults results =
    LimelightHelpers.getLatestResults("limelight");

for (LimelightHelpers.LimelightTarget_Classifier c : results.targets_Classifier) {
    System.out.printf(
        "Class: %s (ID %d)  confidence=%.2f  zone=%.0f  tx=%.1f  ty=%.1f%n",
        c.className, (int) c.classID, c.confidence, c.zone, c.tx, c.ty
    );
}

Each LimelightTarget_Classifier contains:

Field	Type	Description
`className`	`String`	Human-readable class name
`classID`	`double`	Numeric class ID
`confidence`	`double`	Classification confidence (0–1)
`zone`	`double`	Zone index used for classification
`tx`	`double`	Horizontal offset in degrees
`ty`	`double`	Vertical offset in degrees
`tx_pixels`	`double`	Horizontal offset in pixels
`ty_pixels`	`double`	Vertical offset in pixels

Detector pipeline

The detector pipeline locates individual objects in the frame, returning bounding boxes, class labels, and angular offsets for each detected object.

Quick reads (NetworkTables)

// Primary detected object's class name
String detClass = LimelightHelpers.getDetectorClass("limelight");

// Primary detected object's class index
int detIndex = LimelightHelpers.getDetectorClassIndex("limelight");

Raw detections (fastest — no JSON parse)

Prefer getRawDetections() over getLatestResults() for time-sensitive detection loops. Raw detections are read directly from NetworkTables and avoid the overhead of JSON serialization and parsing on the Limelight and deserialization on the RoboRIO.

LimelightHelpers.RawDetection[] detections =
    LimelightHelpers.getRawDetections("limelight");

for (LimelightHelpers.RawDetection d : detections) {
    System.out.printf(
        "Class %d  txnc=%.1f  tync=%.1f  ta=%.2f%n",
        d.classId, d.txnc, d.tync, d.ta
    );

    // Bounding box corners (in pixels)
    // corner0 = top-left, corner1 = top-right
    // corner2 = bottom-right, corner3 = bottom-left
    System.out.printf(
        "  Corners: (%.0f,%.0f) (%.0f,%.0f) (%.0f,%.0f) (%.0f,%.0f)%n",
        d.corner0_X, d.corner0_Y,
        d.corner1_X, d.corner1_Y,
        d.corner2_X, d.corner2_Y,
        d.corner3_X, d.corner3_Y
    );
}

Each RawDetection contains:

Field	Type	Description
`classId`	`int`	Numeric class ID
`txnc`	`double`	Horizontal offset from principal pixel in degrees
`tync`	`double`	Vertical offset from principal pixel in degrees
`ta`	`double`	Target area as % of image
`corner0_X` / `corner0_Y`	`double`	Top-left corner coordinates
`corner1_X` / `corner1_Y`	`double`	Top-right corner coordinates
`corner2_X` / `corner2_Y`	`double`	Bottom-right corner coordinates
`corner3_X` / `corner3_Y`	`double`	Bottom-left corner coordinates

Full results via JSON

LimelightHelpers.LimelightResults results =
    LimelightHelpers.getLatestResults("limelight");

for (LimelightHelpers.LimelightTarget_Detector d : results.targets_Detector) {
    System.out.printf(
        "Class: %s (ID %d)  conf=%.2f  ta=%.2f  tx=%.1f  ty=%.1f%n",
        d.className, (int) d.classID, d.confidence,
        d.ta, d.tx, d.ty
    );
}

Each LimelightTarget_Detector contains:

Field	Description
`className`	Human-readable class name
`classID`	Numeric class ID
`confidence`	Detection confidence (0–1)
`ta`	Target area as % of image
`tx` / `ty`	Angular offset from crosshair in degrees
`tx_pixels` / `ty_pixels`	Offset in pixels
`tx_nocrosshair` / `ty_nocrosshair`	Offset from principal pixel in degrees

Python scripting

You can run custom Python scripts on the Limelight and exchange data arrays over NetworkTables. This lets you implement advanced logic or custom processing without modifying robot code.

// Send data to the Python script running on the Limelight
// Up to 32 doubles in the outgoing array
double[] outgoing = { 1.0, 2.0, 3.0 };
LimelightHelpers.setPythonScriptData("limelight", outgoing);

// Read output from the Python script
double[] incoming = LimelightHelpers.getPythonScriptData("limelight");
if (incoming.length > 0) {
    System.out.println("Python output[0]: " + incoming[0]);
}

Data flows as double[] arrays over the llrobot (outgoing) and llpython (incoming) NetworkTables keys. The Python script reads from llrobot and writes to llpython.

Get Started

Core Concepts

Guides

Classifier pipeline

Quick reads (NetworkTables)

Full results via JSON

Detector pipeline

Quick reads (NetworkTables)

Raw detections (fastest — no JSON parse)

Full results via JSON

Python scripting

Build docs developers (and LLMs) love

Get Started

Core Concepts

Guides

​Classifier pipeline

​Quick reads (NetworkTables)

​Full results via JSON

​Detector pipeline

​Quick reads (NetworkTables)

​Raw detections (fastest — no JSON parse)

​Full results via JSON

​Python scripting

Build docs developers (and LLMs) love

Classifier pipeline

Quick reads (NetworkTables)

Full results via JSON

Detector pipeline

Quick reads (NetworkTables)

Raw detections (fastest — no JSON parse)

Full results via JSON

Python scripting