LEARNING MOTION PARAMETERS FOR A MEJDOOL-DATE THINNING ROBOT

Fruitlet thinning is an important task of Mejdool-date cultivation. Attaining high yields and qualities requires accurate and timely monitored thinning. Currently, thinning is a manual, labor-intensive process. It is estimated that 3.5 hours are required to thin a single mature tree and more than a million hours of labor are currently required to thin all the Medjool trees in Israel. Automation of this process is imperative to enable fruit thinning of larger plots with reduced man power. The Mejdool-date fruit cluster grows within the date crown at the top of the tree, between the palm leaves. Therefore, thinning requires inside-the-canopy operation, hence careful operation is needed to manipulate the thinning process within the date palm crown.

As part of a project for developing a robotic system for Mejdool-date thinning, we have devised a motion planning and control algorithm. The algorithm is based on dynamic movement primitives (DMP). DMPs are units of action that are formalized as stable nonlinear attractor systems. DMPs have parameters whose values can be tuned to facilitate performance of new tasks. The parameters of the DMP equations determine the precision of the generated movement. DMP parameters can be divided into three categories: shape parameters (defining motion trajectory), meta parameters (determine the DMPs frame of operation), and external parameters (determine external, high level, environmental constraints that affect DMP operation).

We propose an innovative method for hierarchical adaptation of DMP parameters during run-time. The adaptation improves the parameter generalization and the precision of the generated movement, which is essential for the thinning operation. To this end, a mapping of the meta parameters as a function of the task parameters is learned a-priori. Since the mapping is a multi-input-multi-output learning problem, it can be modeled using a neural-network, in which the task parameters form the input layer and the meta parameters are the output layer.

To attain good accuracy, the neural-network requires a large dataset. As the relevant period for collecting the data is limited to a few weeks a year and entails long working epochs, establishing such a dataset is challenging. To address this challenge, a validated, stochastic model of a fruit cluster was created and visualized in 3D using python OpenGL. The fruit cluster is modeled as an assembly of geometric shapes, each shape has a unique distribution fitted to it. The model was validated by date-palm experts. The validated, stochastic model enables creation and visualization of an infinite number of fruit clusters facilitating the generation of a large training database required for learning. In an upcoming experiment, we will train a neural-network for mapping task requirements to meta parameters using the model. We shall than test motion based on DMPs adapted using the trained neural-network in a physical laboratory environment.

Keywords: Dynamic motion primitives (DMP), robotics in agriculture, neural networks

Acknowledgments: Research is supported by the Israeli ministry of agriculture and the Israeli Date Grower`s board in The Plant Council.