In recent industrial fields, robot arms are used for various purposes. Robots are being used in many fields, from simple problems such as screwing, to moving objects precisely, or even very delicate tasks required in the medical field. A lot of resear...
In recent industrial fields, robot arms are used for various purposes. Robots are being used in many fields, from simple problems such as screwing, to moving objects precisely, or even very delicate tasks required in the medical field. A lot of research has been done so that these robots can be automatically adjusted, not humans, through deep learning and reinforcement learning. These studies considerably relieve the trouble of having to adjust each robot by humans.
However, the biggest problem with how to operate a robot or robot arm with deep learning and reinforcement learning is that it requires a very large amount of data for learning. There is also a method of learning with a real robot or simulation without a large amount of data, but there is a problem that when learning with a real robot, the real robot may break down or people may be injured. Learning through simulation can also cause problems in actual operation if it is a simulation that does not match the real thing completely, and there is a problem that it costs a lot to create a simulation that matches the real thing perfectly.
This study proposes a model to solve this problem. First, we explore how models such as LAPO, LAPA, and Dynamo constructed the model to solve this problem. Similar to this model, we propose a model that uses a dynamic model to extract information important to behavioral decisions from the model for actual behavior. This makes meaningful learning possible even in situations where there are very few data with 50 or less available behavioral information.