Autonomous Root Tip Inoculation System

Overview

Manual pipetting in plant phenotyping is slow, inconsistent, and cannot keep up with high-throughput greenhouse environments. Existing software was semi-supervised — it still required a human in the loop. The goal was to build a fully autonomous, 24/7 inoculation system for NPEC (Netherlands Plant Eco-Phenotyping Centre).

Pipeline

1. Preprocessing — Raw high-resolution images are cropped, normalized, and split into 128px patches.

2. Mask Prediction — A dual U-Net model predicts binary segmentation masks of root structures, achieving F1 scores of 0.86 and 0.83.

3. Root Tracing & Tip Extraction — Uses Dijkstra’s algorithm for graph theory tracing, trajectory projection to bridge gaps, neighbor reclaim for crossing roots, and biological constraints to remove condensation artifacts.

4. Controller & System Integration — A dynamic calibration pipeline translates pixel coordinates to real-world millimetre positions in real-time.

Controller	Mean Error	Notes
PID	2.13 mm	High consistency, selected for delivery
SAC (RL)	8.89 mm	High variance, unpredictable wandering

Results

Root tip detection

Autonomous inoculation demo 2

Tech Stack

Python, OpenCV, NumPy
U-Net (TensorFlow/Keras)
Dijkstra’s algorithm for graph tracing
PID controller & Soft Actor-Critic (SAC) RL agent
MuJoCo simulation environment
NPEC greenhouse imaging data

Limitations & Future Work

Currently trained on one plant species only
Assumes positive gravitropism (roots grow downward)
Cannot detect roots occluded behind the shoot
PID and RL controllers would benefit from further tuning on real hardware +++