An Implementation of Reinforcement Learning in Assembly Path Planning based on 3D Point Clouds

Abstract

3D point clouds consisting of a lot of informatively geometric data have been playing critical roles in many applications such as 3D segmentation, polyline annotation for lane tracking, and especially in manufacturing industry. In particular, this paper proposes to apply Reinforcement Learning (RL) to resolve an automated assembly task based on 3D point cloud data. To address this task, the proposed structure is separated into 2 stages including registration stage and assembly path planning stage. Firstly, in the registration stage, one of the objects is matched to an assembled model to determine the transformation between two 3D point clouds by using RANdom Sample Consensus (RANSAC) and Iterative Closet Point (ICP). Secondly, we employ Q-learning method to train a model to make optimal decisions in assemble path planning task. The entire optimized assembly path planning task has been successfully accomplished for typical objects. Finally, the performance of the approach developed in this paper has been validated by experiments.