| Popis: |
Data collection is a major bottleneck in vision based robot grasping and manipulation applications. The availability of data is not always possible in confidential or high-risk areas providing a constraint for conducting experiments that require deep learning models. Therefore, synthetic data can be created in simulation and then transferred to the real wold using Domain Randomization (DR) technique that introduces random variations in the data by modifying the physical parameters like colour, texture, background, lighting and orientation. The aim is to bridge the gap between the simulated and real environment as once the model is transferred to the real world, it sees the data as just another variation of the simulation. The project develops a framework for implementing DR technique on non-primitive shapes and using a synthetic dataset to train machine learning models that have comparable results to the real data. DR has been applied on five industrial parts used in diesel engine assembly through Blender rendering software. 3D CAD models were used in the simulator and modifications in the physical dynamics were applied. Texture synthesis was explored using image-based, procedural and Physically Based Rendering (PBR) techniques. In order to wrap the textures on custom objects, UV unwrapping method was used. The dataset was created in Gazebo using segmentation camera and it was trained on object detection algorithms. Then, the testing was done on bounding box, segmentation and keypoint detection using multiple datasets with single class, multi-class with random textures and multi-class with metallic textures. Distractor objects and lighting conditions were added randomly in the data. The detection and segmentation results showed that the model was able to transfer efficiently in the real world setting. However, some fuel lines showed false detections due to their similar curvature. It was found that the orientation of objects and illumination played a critical role in DR transfer to the real world. The study illustrates that it is fast and beneficial to train neural networks on entirely synthetic data and use the machine vision for automating key industrial processes. |
