Skip to main content

Vision Template

The vision template project provides a starting point for creating a high-performance vision or pose estimation system compatible with AdvantageKit's deterministic log replay. It includes support for the following features:

  • Integration with both Limelight and PhotonVision
  • Vision simulation via PhotonLib
  • Options for both simple targeting and full pose estimation
  • High-frequency sampling to ensure that every observation is processed (and never duplicated)
  • Efficient logging of observations via structs
  • Advanced filtering options, including automatic scaling of standard deviations
  • Detailed logging of filters, allowing for easy tuning in replay
  • Deterministic replay with a guarantee of accuracy
info

The AdvantageKit vision template is open-source and fully customizable:

  • No black boxes: Users can view and adjust all layers of the vision processing stack.
  • Customizable: IO implementations can be adjusted to support any hardware configuration.
  • Replayable: Every aspect of the vision processing and filtering logic can be replayed and logged in simulation using AdvantageKit's deterministic replay features with guaranteed accuracy.

⚠️ Warning​

This project is provided as a starting point that will work reasonably well across a variety of situations, but must be customized to fit your specific needs. It is intended as a platform on top of which more optimized systems can be designed. The best pose estimation systems account for a wide variety of factors, including:

  • Game design
  • Robot design
  • Strategic objectives
  • Field tolerances
  • ...and more!

High-quality pose estimation requires frequent iteration of all aspects of the control stack to address these factors, including cameras, mounts, coprocessors, calibrations, pipelines, communication, and filtering. This project provides a starting point for only one small part of a well-optimized vision stack.

Configuration​

The project is primarily configured via the VisionConstants class, with comments explaining the purpose of each field. The selected vision implementation can be changed in the constructor of RobotContainer.

tip

In addition to pose estimation, this project include an example of simple targeting with AprilTags. For many games and robots, this is a significantly simpler method of accomplishing game objectives. Check the getTargetX method of Vision and configureButtonBindings method of RobotContainer for details.

The vision subsystems logs a large set of outputs that can be used for debugging and tuning. Each camera logs the following fields:

  • TagPoses: A list of 3D poses representing the set of visible tags. We recommend visualizing this field using the "Vision Target" object on the 3D field tab in AdvantageScope.
  • RobotPoses: A list of 3D poses representing the raw pose estimates from the last cycle. We recommend visualizing this field using the "Ghost" object on the 3D field tab in AdvantageScope.
  • RobotPosesAccepted: A subset of the RobotPoses list with the set of estimates that passed all stages of filtering.
  • RobotPosesRejected: A subset of the RobotPoses list with the set of estimates that were removed during filtering.

The Summary table includes identical fields which include samples from every camera.