Imitation learning has attracted attention for its ability to learn human manipulation skills and enable robots to adapt to changes in the environment. In particular, bilateral control-based imitation learning has been proven to be effective in tasks that require force adjustment. However, conventional methods do not consider the relationship between the angle and angular velocity of the robot in training the neural network. Robots with inadequate learning of physical relationships may lead to low task success rates and poor generalizability of their movements. In this study, we proposed a learning method that considers the relationship between angles and angular velocities as a loss function in bilateral control-based imitation learning. In experiments, two tasks were conducted by the conventional and the proposed methods, and the effectiveness was verified.
