Learning-Based Picking and Placing Configuration Sampler for Mobile Crane Lift Path Planning

Abstract

Picking and placing configurations (PPCs) are critical for lifting-path planning, but are challenging to determine. The crane location regions (CLRs)-based sampler is an efficient method that can calculate PPCs, but it cannot guarantee the optimality of the generated configurations because it does not take into account the relationship between the distribution of surrounding obstacles in the environment and the position of the crane. To address the problem, we present a novel learning-based optimal configuration sampler, called LOC-Sampler. The sampler is constructed in two stages: offline training of the crane location distribution model based on a neural network, and online generation of an optimal configuration. In the first stage, numerous high-quality crane locations in various lifting tasks are used to train the crane location distribution model. The trained model is used to predict some promising regions that are likely to contain optimal PPCs for a given pose of the lifted object. These promising regions serve as a nonuniform sampling heuristic to limit the number of PPCs in collision-free regions, and are useful for successful planning. The main task of the second stage is to construct a sampler based on the trained crane location distribution model. The sampler then operates on these promising regions to improve the success rate for the lifting path planning and the quality of the path. The results of experiments showed that the proposed method can generate about 50% more feasible PPCs than the CLR-based and inverse kinematics (IK)-based methods. In addition, the PPCs generated by the proposed method usually are superior, resulting in a success rate about 74% higher than that of CLR-based and IK-based methods, and provide an excellent path for lifting path planning. Furthermore, compared with the CLRs-based and IK-based methods, the LOC-Sampler method can quickly predict the optimal distribution of crane locations, instead of obtaining only a specific configuration of a crane according to the given poses of the lifted object and the operation scene information. The proposed LOC-Sampler also can be incorporated into different path planning algorithms, which effectively can improve the efficiency and reliability of path planning, and it provides a competitive and effective PPC, compared with the commonly used CLRs- and IK-based methods, for lifting path planning. The application of crane lifting path planning in construction, manufacturing, and other fields improves engineering efficiency. The crane location region method proposed previously can calculate the configuration quickly to finish the lifting task. However, that crane location region method does not take into account the key factor of the quality of the crane configuration. This research introduces the LOC-Sampler, a machine learning–based tool that optimizes crane path planning. By predicting the best crane positions considering the object’s pose and obstacle layout, LOC-Sampler significantly improves upon traditional methods, increasing path planning success rates by 74%. This innovation is particularly valuable in an industry that is moving toward automation and digital integration, providing a practical solution that can be readily adopted into existing workflows. The LOC-Sampler is designed to be scalable, aligning with the industry’s shift toward smart construction practices. With a successful trial in a three-dimensional (3D) lifting simulation, our tool is poised to offer tangible benefits, making complex crane operations more efficient and secure. This advancement is a step toward the future of construction, in which intelligent systems such as LOC-Sampler will streamline project execution.