In the production process of automotive parts, the 6D pose recognition for unordered bin picking is a crucial and challenging task. WeLinkirt provides a reliable solution for this scenario with its advanced AI vision technology.
User Scenario: An assembly line of a leading automotive parts manufacturer needs to accurately pick specific parts from bins with randomly placed parts for subsequent assembly. The products are various automotive parts with complex shapes, and the detection object is the 6D pose (including position and orientation information) of these parts in the bins.
Pain Points: Traditional picking methods face many quantitative challenges. The miss rate is relatively high, about 6%, which causes some parts to fail to be accurately picked and affects production efficiency. The false alarm rate reaches 8%, making the robotic arm frequently perform invalid picking actions, increasing energy consumption and equipment wear. The labor cost remains high as manual intervention is required to handle failed pickings. Additionally, the changeover time is long. It takes about 30 minutes to re-adjust each time the part type is changed.
Technical Principle
WeLinkirt uses advanced AI algorithms and 3D imaging technology to solve this problem. In terms of algorithms, deep-learning algorithms are used to train a large amount of part image data, enabling the model to learn the features of parts and their performance in different poses. The convolutional neural network (CNN) is used to extract feature information from images, and the recurrent neural network (RNN) is combined to process sequential data, which can better understand the spatial structure and pose relationship of parts.
- In terms of imaging, the self-developed 3D camera can quickly and accurately obtain the 3D morphology information of parts. Its principle is to project a specific pattern onto the part surface through structured light projection technology. The camera captures the deformation of the reflected pattern and calculates the 3D coordinates of each point on the object surface using the triangulation principle, thus reconstructing the 3D model of the part.
- This 3D imaging method can clearly present the details and spatial positions of parts, providing accurate data for subsequent 6D pose recognition.
- At the same time, combined with AI algorithms to analyze the 3D model, it can accurately identify the position and orientation of parts. Even in the case of parts being mutually occluded and randomly placed, high-precision 6D pose recognition can be achieved.
WeLinkirt's Solution and Product Introduction
WeLinkirt uses the DaoAI robot vision product to solve this scenario problem. This product has the capabilities of 6D pose recognition, unordered bin picking (bin picking), gluing/assembly guidance, and a closed-loop of brain, eye, and body. In the implementation process, first, the self-developed 3D camera scans the parts in the bin to obtain their 3D morphology data. Then, the data is transmitted to the DaoAI AI AOI software system. Based on the feature recognition of the visual basic model, this system can quickly and accurately identify the 6D pose of parts. Even with only 1-20 good samples, it can learn and recognize efficiently through the APDT positive-sample/few-sample learning technology, and at the same time, use the semantic false-alarm filtering function to reduce false alarms. Finally, the recognition result is fed back to the robot to guide its accurate picking.
WeLinkirt's DaoAI robot vision product provides a precise solution for unordered bin picking of automotive parts.
Quantitative Results: By applying WeLinkirt's solution, remarkable results have been achieved. The miss rate has been reduced from 6% to <0.6%, greatly improving the part-picking success rate. The false alarm rate has been reduced by -63%, reducing the invalid actions of the robotic arm and lowering energy consumption and equipment wear. The changeover time has been shortened from 30 minutes to 5 minutes, improving production flexibility and efficiency.