AprilTags and Vision Coprocessors

Since 2023, FRC fields carry AprilTags — printed fiducial markers that act like QR codes for robots. They are the foundation of modern on-field localization.

What an AprilTag gives you#

Each tag has a unique ID and a known, published location and orientation on the field (provided each season as the WPILib AprilTag field layout, an AprilTagFieldLayout you can load in code). When a camera detects a tag, it can compute the 3D transform from the camera to the tag. Because the tag's field position is known, the robot can work backward to compute its own field-relative pose. FRC used the 16h5 tag family in 2023, then switched to 36h11 starting in 2024 (and 2025+) because 36h11 substantially reduces false positives and is more widely compatible.

Why a coprocessor#

Detecting tags and doing the pose math is computationally heavy and you do not want to bog down the roboRIO. So vision usually runs on a coprocessor — a small dedicated computer with a camera — that does the image processing and sends results to the roboRIO over the network via NetworkTables. The two dominant options:

• Limelight: an all-in-one smart camera with a polished web UI, on-board AprilTag pipelines, and (on newer models) a built-in IMU. Easiest to set up.
• PhotonVision: free, open-source vision software you run on a coprocessor like an Orange Pi or Raspberry Pi (or a supported camera). Highly flexible, with the PhotonLib vendor library for robot code.

Camera and tuning basics#

For reliable tag detection you need:

• Calibration: the camera's lens parameters must be calibrated so the 3D math is accurate. Both Limelight and PhotonVision provide calibration workflows; do this per resolution.
• Exposure and resolution: lower exposure and a fixed gain reduce motion blur and make tags crisp; higher resolution sees tags farther but costs frame rate. Tune for low latency.
• Robot-to-camera transform: you must tell your code exactly where the camera sits on the robot (position and angle), or the computed pose will be offset.

Latency matters#

Vision results are always a little old by the time robot code uses them, because of capture, processing, and network time. Good vision libraries report a timestamp with each measurement so the pose estimator (next lesson) can account for the delay — never just treat a vision pose as 'now.'