Pek Automotive, Apr 2025 to present
Camera-to-arm calibration for a harvesting robot
A three-stage calibration replaced a manual procedure and reaches 5 mm max positional error (2 mm std) against a 10 mm production threshold.
5 mm max error
Problem
The harvesting robot has several arms and finds fruit with a camera. The camera reports a fruit’s position in its own 3D coordinates; each arm needs that position in its own frame, within a 10 mm production threshold. Until then, the mapping was calibrated by a manual procedure.
Approach
I designed the calibration as three stages, each correcting a different source of error:
- Per-arm 2D polynomial regression maps camera coordinates to that arm’s positions.
- A linear misalignment correction removes the offset that remains between camera and arm.
- Camera tilt compensation corrects for the camera tilting while the arm extends.
Around the fit, the routine rejects outliers, validates the result with R², and retries automatically. It reports its status over UART, so operators can verify a calibration without connecting a PC.
Result
5 mm maximum positional error with a 2 mm standard deviation, against the 10 mm production threshold. The manual procedure is gone.