AprilTag Vision: Knowing Where You Are with PhotonVision

AprilTags are the fiducial markers placed around the FRC field. By detecting them with a camera, your robot can compute its own position on the field — enabling automated alignment, reliable autos, and shot-aiming. The 2026 REBUILT field uses the 36h11 tag family (IDs 1-32), with 8.125-inch tags on the Hub, Tower Wall, Outpost, and Trenches.

The architecture. Image processing is too heavy for the roboRIO, so it runs on a coprocessor — commonly an Orange Pi or Raspberry Pi running PhotonVision — connected by Ethernet to the radio. The camera (e.g. an Arducam OV9281 global-shutter or a Limelight) feeds PhotonVision, which detects tags and publishes results over NetworkTables. Your robot code reads them with the PhotonLib vendor library.

Calibrate first — this is non-negotiable for 3D. To get a tag's 3D pose you must calibrate the camera at the resolution you will run. In PhotonVision you photograph a ChArUco/chessboard board from many angles and distances. Calibration solves for focal length, optical center, and distortion. Skipping or rushing this is a top cause of garbage pose data.

Use the right field layout. For 2026 there are two official field layouts — the welded layout (which PhotonVision ships with) and the AndyMark layout. Select the one matching the event field, both on the coprocessor and in your pose-estimation setup, or every pose will be subtly wrong (the AndyMark vs welded difference can be ~0.5 inch).

MultiTag and pose estimation. Enabling MultiTag estimation lets PhotonVision combine all visible tags using the field-layout JSON to produce one robust camera-to-field pose. In code, PhotonPoseEstimator turns pipeline results into a robot pose. The professional move is to fuse vision into a SwerveDrivePoseEstimator (or DifferentialDrivePoseEstimator) with addVisionMeasurement(). Conceptually:

// Pull unread results, run the estimator, push valid poses in.
for (var result : camera.getAllUnreadResults()) {
  Optional<EstimatedRobotPose> est = photonEstimator.update(result);
  est.ifPresent(e ->
      poseEstimator.addVisionMeasurement(
          e.estimatedPose.toPose2d(), e.timestampSeconds));
}

The exact PhotonLib API moves between seasons — recent versions deprecated getLatestResult() in favor of getAllUnreadResults() and added strategy-specific estimate methods — so check the current PhotonVision docs for your installed version. The fusion idea is stable: blend fast wheel/gyro odometry with absolute vision corrections, so odometry gives smooth high-rate position while vision periodically snaps out accumulated drift. Tune the trust (standard deviations) so noisy single-tag readings don't yank the pose around.

The REBUILT payoff: a robot that knows its field pose can auto-align to the Hub for consistent Fuel shots and run repeatable autos that don't drift over the 20-second period. Create one PhotonPoseEstimator per camera; multiple cameras give more frequent corrections and fewer blind spots.