Original papers
Collision-free motion planning for the litchi-picking robot

https://doi.org/10.1016/j.compag.2021.106151Get rights and content

Highlights

  • Aims at gaining a collision-free motion algorithm for litchi-picking robot.

  • A collision detection algorithm for picking-robots is established.

  • An inverse kinematics solving algorithm base on adaptive weighted PSO algorithm is proposed.

  • An improved Bi-RRT algorithm is proposed, combining with target gravity concept and adaptive coefficient adjustment method.

Abstract

In unstructured environments, picking robots could collide with branches, thus reducing the success rate of picking. In this study, collision-free motion planning is performed in two steps based on the binocular stereo vision information of the environment. First, according to the 3D information collected in the stereo vision environment, the improved adaptive weight particle swarm optimization (APSO) algorithm is adopted to solve the inverse kinematics problem of robots and obtain the collision-free picking posture. Second, to address the limitations of the Bi-RRT algorithm in a high-dimensional environment, such as randomness and slow convergence speed, the target gravity concept and adaptive coefficient adjustment method are introduced to the Bi-RRT algorithm, which is called the AtBi-RRT algorithm. The AtBi-RRT algorithm is used to determine the collision-free path of the robot. Simulation results indicate that the APSO algorithm can swiftly obtain the proper collision-free picking posture, and the average path-determination time of the AtBi-RRT algorithm is 4.24 s. The success rate of path determination is 100% for the laboratory’s picking scene. Experimental results verify that the proposed collision-free motion-planning method allows the picking robot to avoid obstacles in the workspace and efficiently complete the picking task.

Introduction

Fruit picking is a complex task, and the labor cost for it usually accounts for 50% to 70% of the production cost (Tang, 2021). Owing to increasing labor costs and limited available labor, the production cost is increasing; hence, there is a crucial need for picking robot to mechanically perform and ease these production operations. An agricultural robot can be defined as an integration of sensing, computing, manipulation, and control techniques to execute plant production tasks with artificial intelligence. In other words, the successful adoption of these techniques requires integration between the robot’s abilities and its working environment. However, the fruit picking environment is complicated, as the picking robot and end effector are likely to collide with branches obstacles, thus reducing the fruit-picking efficiency, thus, the adoption of robotics in agriculture is faced with considerable challenges. To successfully adopt robots in agriculture, it is necessary to develop an efficient collision-free motion-planning algorithm for agricultural robots to complete picking tasks.

Agriculture is a constrained dynamic environment, and work objects vary in shape, size, position, and orientation. The robot’s target pose is unknown in each picking task, the path from the starting point to the target point of each picking task needs to be replanned. Furthermore, to meet the picking efficiency of the picking robot, the calculation speed of the motion-planning algorithm should be high. Therefore, to meet the above requirements, a collision-free motion-planning algorithm of picking robots is not only required to solve the inverse kinematics problem and obtain the collision-free picking posture with few iterations based on environmental information, but also to swiftly determine a picking path to avoid collisions between the picking robot and the branch obstacle.

Many researchers have proposed many algorithms for inverse kinematics posture and collision-free path determination of robots during the last few decades. In the research on inverse kinematics posture determination of robots, Van Henten designed a general method to solve the inverse kinematics posture of the picking robot. The study was re-formulated as a nonlinear programming problem, and the genetic algorithm was used to solve the problem (Van Henten, et al., 2010). Sun-Oh Park proposes an algorithm for combining the Jacobian-based numerical approach with a modified potential field to solve real-time inverse kinematics for redundant robots in unknown environments (Park, S. O., et al., 2020). Ye H. proposed that different motions can be decoupled by redundant Euler angles represented by four rotations, after which, a closed solution of the inverse kinematics can be determined (Ye H., et al., 2020). Lopez-Franco Carlos proposed soft computing methods to determine the position and orientation of the inverse kinematics problem (Lopez-Franco C., et al., 2018). In the research on the collision-free path determination of robots, the artificial potential field method is effective in addressing the challenge of path planning with low degrees of freedom. The robot obstacle avoidance method based on the artificial potential field was first proposed by Khatib, and it remains the basis of many motion optimization methods (Lazarowska, 2019, Sudhakara et al., 2018, Liu et al., 2017). Moreover, LaValle proposed a rapidly-exploring random tree algorithm (RRT) based on random sampling in the configuration space to determine a path. The RRT algorithm is effective in path planning for robots, especially for robots with a high degree of freedom. Researchers have proposed many optimization algorithms based on the RRT algorithm. To reduce the computation time, Kuffner and LaValle proposed the Bi-RRT algorithm, which grew two trees from the initial and the target states simultaneously, thus improving the exploration and convergence speeds of the algorithm (Lavalle S.M., et al., 2000). Bertram et al. introduced a heuristic growth method. The basic principle is to define an objective function, which guides the growth of the search tree along with the target point (Bertram D. et al., 2006).

There have been several studies and applications in the solution of the inverse kinematics and path planning of robots. However, these studies focused on industrial robots with uniform objects in an unconstrained workspace. Only a few studies have reported collision-free path planning for picking robots. Bac et al. conducted path planning research on constrained-azimuth picking robots in intensive planting environments; they proposed a constrained-azimuth path planning method (Bac C.W. et al., 2016). Cao et al. adopted genetic algorithms and artificial potential field ideas to optimize the RRT algorithm and designed a path planning algorithm applied to litchi-picking robots; however, in scenarios with many branches and obstacles, the planning time increases, whereas efficiency decreases (Cao X., et al, 2019).

This study proposed a collision-free motion-planning method for a litchi-picking robot based on the studies mentioned above. First, a 6-degree of freedom (DOF) robot was adopted as the picking robot, a binocular vision was used to perceive the environment, and a self-developed binocular vision software was utilized to identify and locate the target (Chen et al., 2020, Lin et al., 2021, Luo et al., 2016, Wang et al., 2016, Zou et al., 2016). Second, to reduce the computational complexity, this study practically simplified the robot arm and obstacle models to construct an obstacle collision-detection model. Third, an improved adaptive weighted particle swarm optimization algorithm (APSO) was adopted to solve inverse kinematics problems and obtain a collision-free picking posture for the picking robot in a complex picking environment. Fourth, to accelerate the path-determination speed, the AtBi-RRT algorithm was proposed, which is based on the Bi-RRT algorithm and introduces the target gravity concept and adaptive parameter adjustment method. Then, the smoothing method was adopted to optimize the path generated by the AtBi-RRT algorithm, such that an optimal or approximately optimal path can be obtained. Finally, the collision-free motion-planning algorithm was applied to the simulation, and the litchi-picking experiments were used to verify the algorithm.

Section snippets

Kinematic model of the litchi-picking robot

Fig. 1 illustrates the structure of the picking robot. The end of Joint 6 was equipped with an end effector to pick the target fruit. In addition, a binocular vision system was installed on the end effector to obtain 3D point cloud data of the picking environment.

In practical applications, to avoid interference between the front of the end effector and obstacles during the robot movement, as well as compensate for the visual error, an auxiliary target picking point needs to be set 100 mm in

Experiments and results

To investigate the performance of the collision-free motion planning, which includes the APSO and AtBi-RRT algorithms in the picking scene, we set up 10 real picking scenes for simulations and experiments. The scene data obtained using the binocular vision system are presented in Table 2.

Conclusion

  • (1)

    This study adopts the APSO algorithm, which combines with the fitness function suitable for picking the manipulator to solve the inverse kinematics of the picking robot and obtain the collision-free picking posture of the robot. The positioning accuracy and orientation accuracy of the APSO algorithm are more significantly improved than the PSO algorithm.

  • (2)

    Based on the Bi-RRT algorithm, this study introduced the target gravity concept and adaptive adjustment parameter method to the Bi-RRT

CRediT authorship contribution statement

Lei Ye: Conceptualization, Methodology, Software, Writing - original draft. Jieli Duan: Data curation, Project administration, Software. Zhou Yang: Supervision, Resources. Xiangjun Zou: Supervision, Funding acquisition, Writing - review & editing. Mingyou Chen: Visualization, Investigation. Sheng Zhang: Resources.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgement

This work was supported by the Science and Technology Planning Project of Guangdong Province (2019A050510035), the Major scientific research projects of Guangdong Province (No. 2020KZDZX1037). We also thank the anonymous reviewers for their critical comments and suggestions for improving the manuscript.

References (23)

Cited by (36)

View all citing articles on Scopus
View full text