One of the most significant challenges of 21st-century robotics is to develop robots capable of acquiring progressive knowledge and skills autonomously and safely. These flexible robotic systems are expected to revolutionize society by their ability to perform a wide range of tasks in dynamic, unstructured, unpredictable, and/or dangerous environments. They can impact various fields such as industry, aerospace, transport, e-commerce, healthcare, medicine, etc. For instance, they will be able to replace or assist a human in hazardous industrial areas, explore unfamiliar environments, help the elderly and/or disabled in their daily tasks, assist a surgeon in delicate operations, take part in relief efforts in the event of climatic disasters, etc.
Learning behavioral skills is the domain of reinforcement learning algorithms, which acquire behaviors through trial and error. These deep learning systems are difficult to analyze. Indeed, proving (for example) the convergence of deep learning algorithms such as backpropagation is a difficult task in a highly non-convex, high-dimensional problem. In this context, theoretical analysis of stability, convergence, and robustness is an essential procedure for ensuring that robotic systems operate predictably and safely. Furthermore, keeping a human-level interpretability and explainability of the decision-making process of machine learning policies is another challenging objective in end-to-end learning process. Advancing in these fields is a necessary step for building industry-grade standards for robot control.
Therefore, this Research Topic gives priority to research focusing on knowledge and results that can improve the effectiveness of learning methods, especially targeted to meeting the requirements of actual deployment, regardless of the field of application. We especially encourage Original Research, Systematic Reviews, Data Reports on the following themes (but not limited to):
- End-to-end control/model architectures
- Deep learning architectures and algorithms suitable for embedded systems
- Deep neural embedded systems
- Theory of neural networks for convergence, robustness, and other characteristics
- Deep reinforcement learning control of robotic systems
- Dynamic neural networks (DNN) to control robots
- Innovative design to improve the performance of End-to-end learning (convergence, robustness, stability solving speed, control performance, learning efficiency),
One of the most significant challenges of 21st-century robotics is to develop robots capable of acquiring progressive knowledge and skills autonomously and safely. These flexible robotic systems are expected to revolutionize society by their ability to perform a wide range of tasks in dynamic, unstructured, unpredictable, and/or dangerous environments. They can impact various fields such as industry, aerospace, transport, e-commerce, healthcare, medicine, etc. For instance, they will be able to replace or assist a human in hazardous industrial areas, explore unfamiliar environments, help the elderly and/or disabled in their daily tasks, assist a surgeon in delicate operations, take part in relief efforts in the event of climatic disasters, etc.
Learning behavioral skills is the domain of reinforcement learning algorithms, which acquire behaviors through trial and error. These deep learning systems are difficult to analyze. Indeed, proving (for example) the convergence of deep learning algorithms such as backpropagation is a difficult task in a highly non-convex, high-dimensional problem. In this context, theoretical analysis of stability, convergence, and robustness is an essential procedure for ensuring that robotic systems operate predictably and safely. Furthermore, keeping a human-level interpretability and explainability of the decision-making process of machine learning policies is another challenging objective in end-to-end learning process. Advancing in these fields is a necessary step for building industry-grade standards for robot control.
Therefore, this Research Topic gives priority to research focusing on knowledge and results that can improve the effectiveness of learning methods, especially targeted to meeting the requirements of actual deployment, regardless of the field of application. We especially encourage Original Research, Systematic Reviews, Data Reports on the following themes (but not limited to):
- End-to-end control/model architectures
- Deep learning architectures and algorithms suitable for embedded systems
- Deep neural embedded systems
- Theory of neural networks for convergence, robustness, and other characteristics
- Deep reinforcement learning control of robotic systems
- Dynamic neural networks (DNN) to control robots
- Innovative design to improve the performance of End-to-end learning (convergence, robustness, stability solving speed, control performance, learning efficiency),
