Research On 3D Object Detection And Code Generation Technologies For Industrial Robots

With the rapid development of information technology and artificial intelligence,the concept of smart factories and industrial internet has become widely promoted.Robots,as a major component,offer many advantages such as the ability to complete repetitive and monotonous tasks,perform high-risk operations,and improve production efficiency for enterprises.However,the deployment of robot production lines in manufacturing enterprises is often a complex and time-consuming process,particularly in terms of programming and debugging robot code.Accelerating the deployment of robot production lines and realizing quick robot programming is thus crucial for enterprises to improve efficiency and reduce product time-to-market.To address this challenge,this thesis focuses on the technology of robot code generation in industrial environments,and aims to convert natural language instructions into target robot code to achieve this goal.And robot code generation requires information on the position and posture of industrial objects.Accurate positioning coordinates are crucial for robots to successfully complete various tasks.To address these challenges,this thesis introduces 3D object detection technology to facilitate the generation of accurate and safe robot code.Based on the above analysis,this thesis conducts research on 3D object detection and code generation technology in industrial environments.The main contributions of this work are as follows:Firstly,aiming at the problem of industrial object detection in an industrial environment,a 3D object detection method based on an enhanced backbone network and deformable Transformer is proposed.This method describes local features by defining additional geometric information at predefined points in the point cloud,which expands the perception range.Then,the information from each layer of the multilayer perceptron in the backbone network is combined to adaptively enhance the network,and the k-Closest Point Sampling method is used to sample feature points to generate initial candidate objects.Finally,an internal relationship among input points is captured through self-attention,and the candidate objects are used as a reference for adaptive learning of sampling positions.After feature extraction is performed through deformable attention aggregation of key points,the results are fed into a feedforward neural network to obtain the final 3D object detection results.Secondly,aiming at the problem of industrial robot code generation,a robot code generation method based on multi-source information API recommendation is proposed.This method mainly utilizes an API recommendation model to obtain additional information provided by the industrial environment and converts the environmental features into semantic information represented by recommended APIs.Afterwards,API recommendations and location information are embedded into the code generation model.The code generation model uses an encoder-decoder structure based on attention to decompose the abstract syntax tree(AST)into a series of construction actions by following the custom robot code Abstract Syntax Description Language(ASDL),and finally generates the target robot code by converting the AST into machine-executable code via the AST2 CODE function.Experiments on large-scale public datasets verify the effectiveness of the above methods.Compared with the mainstream methods on the real production scenario datasets,the experimental results show that the proposed methods have the best performance.