Robotic grasping of novel objects based on a feature detection algorithm trained on minimal data

Abstract

In recent years, the integration of deep learning into robotic grasping algorithms has allowed for widespread advancements in this field. Most of the current deep learning-based grasping algorithms must be trained on huge amounts of data to deal with a large variety of objects. However, it is inefficient and impossible to train an algorithm to recognize every single object and therefore many existing methods perform poorly when encountering novel objects which are not in the training dataset. This thesis presents a new grasping algorithm utilizing deep learning-based object detector that can deal with novel objects through oriented detection of several key features present in most objects. Combining these features with information extracted via image segmentation, a grasping pose can be logically deduced and hence is not limited by the training size. The proposed algorithm does not require a large amount of training data for feature learning and can have training data more than 100 times less than most other algorithms. The training time of about 10 hours is also significantly lesser compared to some works requiring up to 2 months. Yet, the proposed algorithm is able to achieve a high grasp success rate of 98.25% in actual experiments when encountering a wide variety of novel objects which are not included in the training set. This grasping algorithm can be useful in many applications such as logistics and reconnaissance. This thesis further explores integration of 5G technologies with this algorithm. The high-speed data transfer of 5G over other forms of communication technologies further enable the robot to execute remote grasping of unfamiliar objects in an environment that is risky for humans to enter in real time with low latency.