[Abstract]One of the biggest limitations faced by current robot technology is the difficulty in adapting to changing tasks. When robots face new environments or learn new tasks, they will inevitably experience of old environments or tasks. In order to summarize the research and development status of continuous learning in robots, we first introduced the framework and evaluation benchmarks of continuous learning, then elaborated on the necessity and challenges of continuous learning in robot tasks, and reviewed the development status of continuous learning. Finally, the development prospects of continuous learning of robots are looked forward to, and some valuable research questions are raised.
[Keywords] Continuous learning; lifelong learning; robot
0 Preface
A robot is an intelligent system that combines knowledge in the fields of artificial intelligence and automatic control and can replace humans in operating in complex environments. Thanks to the rapid development of robots and the continuous changes in hardware equipment, robots have achieved outstanding performance in many fields. For example, robot dexterous hands that imitate the dexterity of human hands, mass-produced autonomous vehicles, etc. Any robot can be described as the implementation of a set of mappings from action and perception signals at the previous moment to action and perception signals at the next moment. Similar to humans, robots need to evaluate the merits of strategies and perform optimal actions in highly challenging environments filled with complex objects and complex social environments involving humans, animals, and other robots.
The original goal of autonomous robotics is to build physical systems that can complete specific tasks in real environments without human intervention. Despite considerable advances in robotics over the past decade, this goal has not yet been achieved. One of the larger limitations of current robotics is the lack of adaptability to changing environments and tasks. While some robots have been successfully used in dynamic and complexly perceived environments, their adaptability and fault tolerance are often limited by human predictions of the scenarios the robot may encounter. Since robots need to work in real and potentially unknown environments, it is unrealistic to pre-program every environment a robot may face. The only way to address these challenges is to equip robots with the ability to adapt to their environment and develop appropriate strategies to adapt to potentially changing tasks. In addition, most current machine learning methods are based on offline tasks for experiments, that is, the data of all tasks can be obtained simultaneously. Under this settingÿ