Artificial intelligence (AI) and robotics research have the potential to revolutionize various industries and tasks. However, deploying these technologies in real-world scenarios can be challenging due to issues such as the need for large amounts of annotated data and the difficulty of adapting to different environments and lighting conditions. In this thesis, we present two main contributions towards making it easier to deploy AI and robotics research in real-world scenarios.
Firstly, we propose the use of a shallow network for the detection of the grasp position of an object. Unlike other networks that require a large amount of computation and training time, our shallow network is trained with synthetic images and is able to detect the grasp position efficiently. This is particularly useful for tasks where quick and precise grasp positioning is essential, such as in automated assembly lines or pick-and-place operations.
Secondly, we introduce the use of mobile robots and invisible markers to automatically collect a large amount of annotated data for object detection and grasp position estimation. The invisible markers, which are only visible under UV light, allow us to generate a large amount of real data in a short amount of time and accurately annotate it. This is a significant improvement over synthetic data, which may not generalize well to real-world scenarios, and manual data collection, which is time-consuming and labor-intensive.
In addition to these main contributions, we also discuss the use of key-point estimation for 6D pose estimation and a deep-learning method for dealing with harsh lighting conditions in depth data acquisition. Overall, this thesis presents practical solutions for deploying AI and robotics research in real-world scenarios, enabling the efficient and effective use of these technologies in a variety of applications.