Monitoring crop nutrient levels autonomously can allow farmers to quickly adjust fertilizer usage in their fields to maximize their harvest. Previous research has focused on automating crop inspection through visual tracking of growth, which can delay the detection of nutrient deficiencies as visual symptoms may be slower to develop. Instead, we present a contact-based phenotyping robot platform that can autonomously insert nitrate sensors into corn stalks. This task is challenging not only because of the varying field conditions encountered on the farm but also because inserting sensors requires sub-centimeter precision in an environment that has high clutter and occlusion. To deal with these challenges, we create a custom robot gripper with compliant mechanisms for interacting with the cornstalks and develop a robust perceptionaction pipeline to deploy sensors. Through our experimental validation at a cornfield in Iowa, we demonstrate our robot platform’s capability of autonomously inserting sensors into 48 cornstalks over a two-day trial period. Our research platform is all open-sourced with a detailed appendix.
Our robot platform, Cornbot, is a four-wheel-drive mobile platform attached with a 6 DOF xArm robot arm that has custom end effector for sensor insertion tasks. This mobile base was built upon the commercially available Amiga from Farm-NG.
We create a simple and compact gripper for inserting nitrate sensors. As there are high variation in stalk thickness and high clutter in the field, gripper has clever mechanical design such as funneled edges and springloaded compliance to improve alignment.
The pipeline is built as a ROS Service that returns a stalk 3D position given series of RGB-D frames. In each frame, stalk 2D masks are first extracted, and then a best stalk is determined from the 3D pointcloud. After all frames are processed, a best insertion position is determined by a consensus among frames.
Through our experimental validation at a cornfield in Iowa, we demonstrate our robot platform’s capability of
autonomously inserting sensors into 48 cornstalks over a two-day trial period.
Each trial took about three minutes for the robot to perform and one minute for human evaluators to examine the insertion quality.
To examine the insertion quality, we visually inspected whether the sensor was inserted deep enough to cover the entire
sensor pads, and cut open the stalk after insertion to verify
the sensor was actually in the pith region of the stalk. Failure analysis of the 48 insertion trials is shown below.
