Automated Root Phenotyping & Robotic Inoculation

NPEC studies how plant genetics and environment interact. Its Hades module images tens of thousands of seedlings, but inoculating each root with microbes was still manual: slow, difficult to scale, and variable. My project automated the step end to end in simulation - detect a root, find a biologically meaningful point, and drive a liquid-handling robot to it.

OpenCV first finds and crops the Petri dish. A TensorFlow/Keras U-Net then labels root, shoot, and seed pixels. Because roots occupy only a fraction of one percent of an image, I used weighted cross-entropy plus Dice loss and evaluated with F1 rather than misleading pixel accuracy. Skeletonization traces each primary root from shoot to tip.

For control, I built a three-axis PID controller with anti-windup for a simulated Opentrons OT-2 in PyBullet. I also wrapped the simulator in a custom Gymnasium environment and trained a PPO agent, then benchmarked it against PID on identical targets. A coordinate transform joins the two halves by translating image pixels into reachable robot coordinates.

Root segmentation reached about 87% F1. Root-length extraction achieved 6.4% sMAPE after fixing a plant-position matching error that had kept it near 20%. The integrated PID system achieved 94.4% sub-millimeter success with 0.54 mm mean error - a 52% reduction from baseline - while the classical controller proved more precise and consistent than PPO for this task.

Every reinforcement-learning run was tracked in ClearML, remotely executed, checkpointed, and ranked through a structured hyperparameter search with fixed seeds and a versioned model registry.