Hive Vision

Overview

Hive Vision provides BIOBUZZ object-detection models and a separate Java color-detection implementation. Its v2.0.0 release supplies a TensorFlow Lite model for a Limelight 3A. Start by printing the detected class and image offset while leaving the drivetrain still.

Project status

Hive Vision v2.0.0 was released on 12 October 2026. This guide uses that model release with FTC SDK 12's Limelight driver. Measure detection accuracy and inference speed in your own camera setup.

Setup

Deployment paths

This lesson uses the Limelight path. The model is installed on the camera; it is not a new Java dependency for TeamCode. The older PC YOLO and Lab tracks are archived in the project's current documentation.

Camera setup

1. Configure a Limelight 3A using the manufacturer's FTC wiring and setup procedure. Its documented Control Hub connection uses USB-C to the blue USB 3 port. For initial computer setup, use the camera's Hardware Manager or its USB-connected browser interface.
2. Download the v2.0.0 model asset best_limelight3a_ssd_mobilenetv2_300x300_v2_int8_dets25.tflite and matching labels.txt.
3. In a Neural Detector pipeline, select CPU, upload both files, and use the model's 300 × 300 input size. Put this pipeline in slot 0 for the example below.
4. Check that labels are ordered yellow_pollen, red_nectar, blue_nectar with class IDs 0, 1, and 2.
5. Scan/configure the device on the Robot Controller and name it limelight.

The v2 release suggests an initial detector threshold around 80–90%. Treat that as a camera-UI starting point and test both false detections and missed objects. Older setup pages describe different model names and thresholds; use the release's matching assets.

Do not assume limelight.local:5801 is reachable from every robot Wi-Fi connection. Initial USB computer access and access through the robot network are different connection arrangements.

Example

Save this complete Java OpMode as HiveDetectionPreview.java. It was compiled against FTC SDK 12 and needs no dependency beyond the standard SDK project for this integration.

package org.firstinspires.ftc.teamcode;

import com.qualcomm.robotcore.eventloop.opmode.OpMode;
import com.qualcomm.robotcore.eventloop.opmode.TeleOp;
import com.qualcomm.hardware.limelightvision.Limelight3A;
import com.qualcomm.hardware.limelightvision.LLResult;
import com.qualcomm.hardware.limelightvision.LLResultTypes;

@TeleOp(name = "Hive Detection Preview")
public class HiveDetectionPreview extends OpMode {
    private Limelight3A limelight;

    @Override public void init() {
        limelight = hardwareMap.get(Limelight3A.class, "limelight");
        limelight.setPollRateHz(50);
        limelight.pipelineSwitch(0); // Slot 0 must contain the Hive v2 neural model.
        limelight.start();
    }

    @Override public void loop() {
        LLResult result = limelight.getLatestResult();
        if (result == null || !result.isValid()
                || result.getStaleness() > 120 || result.getPipelineIndex() != 0) {
            telemetry.addLine("No recent valid result from pipeline 0");
        } else {
            telemetry.addData("Camera timestamp", result.getTimestamp());
            telemetry.addData("Detections", result.getDetectorResults().size());
            int index = 0;
            for (LLResultTypes.DetectorResult detection : result.getDetectorResults()) {
                telemetry.addData("Detection " + index++, "%s: tx=%.1f deg, raw confidence=%.3f",
                        detection.getClassName(), detection.getTargetXDegrees(),
                        detection.getConfidence());
            }
        }
        telemetry.update();
    }

    @Override public void stop() {
        if (limelight != null) limelight.stop();
    }
}

The 120 ms check is an illustrative receive-age bound, not a universal camera timeout. Tune a freshness policy from measured camera and robot-loop timing. stop() ends the polling owned by this OpMode.

Testing

Place one object in view and move it left and right. Record the reported label, tx, and confidence. Remove it, then add a similar-colored distractor. Check whether the result disappears and whether the class remains correct across lighting, distance, and partial occlusion.

Limits

Result meaning

tx is a horizontal angular offset in degrees. It is not field X, distance, or a complete target pose. A class label is an observation to validate before using it for acquisition or piece counting.

The SDK's detector-confidence documentation and forwarded firmware value do not establish a consistent scale by themselves. This example prints raw confidence. Inspect actual output before implementing a numerical filter; do not assume that a threshold of 0.8 and one of 80 mean the same thing.

getStaleness() measures time since the Control Hub received/parsed a result. It does not prove the image is a newly captured frame. Polling at 50 Hz can read the same camera frame several times. If requiring several independent detections, compare camera timestamps rather than counting repeated loop iterations. Keep camera-local and Control Hub time domains separate.

Webcam integration

The v2 webcam code needs repairs before it is a copy-and-run example. Its calibration import does not match FTC SDK 12, and its returned annotated Mat is handled as VisionPortal draw context rather than a displayed replacement frame. That clone also needs explicit ownership and release. The rolling setup guide uses the wrong Size type for the VisionPortal builder.

For a webcam example, use Color Blob Locator or custom VisionProcessors. Do not combine the Limelight model setup with the separate HSV processor's imports and lifecycle.