FRC Game Piece Position Estimation¶
A fast, robust, and field-ready algorithm for real-time object pose estimation in FRC.
This project provides a complete, open-source pipeline for estimating the real-world position and orientation of FRC game pieces from camera images. It uses a data-driven image-matching approach, offering extremely fast and accurate pose estimation—optimized for Raspberry Pi deployment in real-world matches.
The algorithm is designed for fixed-camera setups and leverages a precomputed dataset of synthetic images rendered in Blender. During runtime, it matches live camera frames to this dataset to infer the pose of the game piece in under 10 milliseconds.
Key Features¶
-
High-Speed & Accuracy
Matches live camera images to a precomputed dataset of rendered frames for sub-10ms pose estimation. -
Fully Explained, Fully Open
Includes all source code, rendering tools, and clear documentation of each stage—perfect for learning, customization, or direct use. -
Modular, Multi-Process Architecture
Ensures real-time performance by offloading detection, estimation, and networking to separate processes. -
Built for FRC Robots
Seamless integration with FRC systems via NetworkTables, optimized for Raspberry Pi or similar edge devices.
How It Works¶
Preprocessing (One-Time, Offline)¶
- Render thousands of labeled images in Blender using the provided scripts.
- Each frame encodes the 3D pose of the game piece from a fixed camera viewpoint.
- Store the dataset in CSV format with pose metadata.
Runtime (On-Robot)¶
- Capture live frames from the robot’s fixed-position camera.
- Detect the game piece using either simple color filtering or ML-based segmentation.
- Extract bounding rectangles and (optionally) orientation info.
- Find the closest match in the dataset using a fast image similarity search.
- Output the real-world position and orientation of the game piece.
Step-by-Step Setup Guide
For detailed setup instructions, Blender configuration, and usage examples, see the Step-by-Step Guide.