The ever-increasing complexity of robot applications induces the need for methods to approach problems with no (viable) analytical solution. Deep learning (DL) provides a set of tools to address this kind of problems. This survey presents a categorization of the major challenges in robotics that leverage DL technologies and introduces representative examples of successful solutions for the described problems. We also consider the question when and whether to use modular, monolithic models or end-to-end DL, in order to provide a guideline for the selection of the correct model structure and training strategy. By doing so, the current role and adaptability of different techniques at different hierarchical levels of a robot-application can be highlighted, thus providing a well-structured basis to assist future approaches.

Recent problems in robotics can sometimes only be tackled using machine learning technologies, particularly those that utilize deep learning (DL) with transfer learning. Transfer learning takes advantage of pretrained models, which are later fine-tuned using smaller task-specific datasets. The fine-tuned models must be robust against changes in environmental factors such as illumination since, often, there is no guarantee for them to be constant. Although synthetic data for pretraining has been shown to enhance DL model generalization, there is limited research on its application for fine-tuning. One limiting factor is that the generation and annotation of synthetic datasets can be cumbersome and impractical for the purpose of fine-tuning. To address this issue, we propose two methods for automatically generating annotated image datasets for object segmentation, one for real-world and another for synthetic images. We also introduce a novel domain adaptation approach called filling the reality gap (FTRG), which can blend elements from real-world and synthetic scenes in a single image to achieve domain adaptation. We demonstrate through experimentation on a representative robot application that FTRG outperforms other domain adaptation techniques, such as domain randomization or photorealistic synthetic images, in creating robust models. Furthermore, we evaluate the benefits of using synthetic data for fine-tuning in transfer learning and continual learning with experience replay using our proposed methods and FTRG. Our findings indicate that fine-tuning with synthetic data can produce superior results compared to solely using real-world data.

In modern robot applications, there is often a need to manipulate previously unknown objects in an unstructured environment. The field of grasp-planning deals with the task of finding grasps for a given object that can be successfully executed with a robot. The predicted grasps can be evaluated according to certain criteria, such as analytical metrics, similarity to human-provided grasps, or the success rate of physical trials. The quality of a grasp also depends on the task which will be carried out after the grasping is completed. Current task-specific grasp planning approaches mostly use probabilistic methods, which utilize categorical task encoding. We argue that categorical task encoding may not be suitable for complex assembly tasks. This paper proposes a transfer-learning-based approach for task-specific grasp planning for robotic assembly. The proposed method is based on an automated pipeline that quickly and automatically generates a small-scale task-specific synthetic grasp dataset using Graspit! and Blender. This dataset is utilized to fine-tune pre-trained grasp quality convolutional neural networks (GQCNNs). The aim is to train GQCNNs that can predict grasps which do not result in a collision when placing the objects. Consequently, this paper focuses on the geometric feasibility of the predicted grasps and does not consider the dynamic effects. The fine-tuned GQCNNs are evaluated using the Moveit! Task Constructor motion planning framework, which enables the automated inspection of whether the motion planning for a task is feasible given a predicted grasp and, if not, which part of the task is responsible for the failure. Our results suggest that fine-tuning GQCNN models can result in superior grasp-planning performance (0.9 success rate compared to 0.65) in the context of an assembly task. Our method can be used to rapidly attain new task-specific grasp policies for flexible robotic assembly applications.