Automated Behavior Planning for Fruit Tree Pruning via Redundant Robot Manipulators: Addressing the Behavior Planning Challenge

TL;DR

This work addresses automated pruning with high-dimensional, redundant manipulators in collision-rich orchard environments. It presents a holistic pruning framework that runs from labeled point clouds to cutting commands, integrating AdTree-based tree modeling, circle-based pose generation, a diverse IK solver (IKFlow), and a holistic pose–motion planner accelerated by VAMP, with a PBS-based approaching controller. The key contributions include a geometry-driven, multi-level planning approach that exploits Cartesian and joint-space redundancies, the integration of diverse IK solutions to navigate complex constraints, and an extensive real-world evaluation showing improved planning efficacy and actionable insights into remaining limitations. The practical impact lies in enabling more reliable, scalable, and less labor-intensive pruning operations, bridging perception and manipulation in autonomous orchard robotics.

Abstract

Pruning is an essential agricultural practice for orchards. Proper pruning can promote healthier growth and optimize fruit production throughout the orchard's lifespan. Robot manipulators have been developed as an automated solution for this repetitive task, which typically requires seasonal labor with specialized skills. While previous research has primarily focused on the challenges of perception, the complexities of manipulation are often overlooked. These challenges involve planning and control in both joint and Cartesian spaces to guide the end-effector through intricate, obstructive branches. Our work addresses the behavior planning challenge for a robotic pruning system, which entails a multi-level planning problem in environments with complex collisions. In this paper, we formulate the planning problem for a high-dimensional robotic arm in a pruning scenario, investigate the system's intrinsic redundancies, and propose a comprehensive pruning workflow that integrates perception, modeling, and holistic planning. In our experiments, we demonstrate that more comprehensive planning methods can significantly enhance the performance of the robotic manipulator. Finally, we implement the proposed workflow on a real-world robot. As a result, this work complements previous efforts on robotic pruning and motivates future research and development in planning for pruning applications.

Figures (12)

• Figure 1: The challenge of the pruning manipulation. (a) An apple orchard during the pruning season presents a unique challenge for robotic pruning. Despite the large-scale organization of the orchard, the complex arrangement of branches creates significant collision issues. (b) Human operators achieve a high success rate by skillfully selecting approaching poses that comprehensively consider collisions and accessibility. (c) In order to tackle the intricate collision during pruning, the robot needs to explore the redundancy in both Cartesian and joint space. The green color indicates the valid solutions while the red color indicates collisions.
• Figure 2: The workflow of our system consists of multiple modules. The system's input is the point cloud data captured by a depth camera. The following procedures will be conducted: (1) Coordinate Transformation: The observation data needs to be transformed from the camera frame to the robot's frame to align perception and robotic behavior. (2) Segmentation: The observed data tend to contain environmental noise such as ground and outliers. (3) AdTree Modeling: Generate an AdTree from the point cloud data. (4) Cut Defining: With labeled point cloud data, we can define cutting locations on the AdTree. (5) Pose Generating: For each cutting location, the planner will generate multiple potential cutting poses to approach the cutting position and choose one based on accessibility. (6) IKFlow: A learning-based IK solver, it can provide diverse IK solutions for one end-effector pose. (7) Holistic Planner: A planner that comprehensively plans the end-effector pose, joint configuration, and trajectory.
• Figure 3: (a) Labeled point cloud input: Our method takes a labeled point cloud as input. In the example shown, red points represent branches to be removed, while blue points represent parts to be kept. (b) To identify cutting locations in the labeled point cloud on the AdTree, we first assign labels to the AdTree's vertices. A cutting location is detected on an AdTree vertex when a consecutive point has a different label. In the figure, the gray cylinders represent sections of the AdTree, and the large dots inside the cylinders represent vertices, with their colors indicating the labels.
• Figure 4: The visualization illustrates the AdTree. AdTree abstracts a tree model from the point cloud. Its topological structure is represented as a geometric tree in the simulation. In the figure, cylinders are generated based on vertex, edge, and thickness data, visually representing different parts of the tree. The blue dots represent the point cloud.
• Figure 5: In the first stage, the robot moves from its initial pose to an approaching pose using a collision-free motion plan. In the second stage, the robot advances from the approaching pose to the cutting pose, positioning the target branch between the shears before making the cut. The approaching pose is shown from both a wide view and a close-up perspective.