Abstract & Overview
In the human–robot interaction, especially when hand contact appears directly on the robot arm, the dynamics of the human arm presents an essential component in human–robot interaction and object manipulation. Modeling and estimation of the human arm dynamics show great potential for achieving more natural and safer interaction. To enrich the dexterity and guarantee the accuracy of the manipulation, mapping the motor functionality of muscle using biosignals becomes a popular topic. In this article, a novel algorithm was constructed using deep learning to explore the potential model between surface electromyography (sEMG) signals of the human arm and interaction force for human–robot interaction. Its features were extracted by adopting the convolutional neural network from the sEMG signals automatically without using prior knowledge of the biomechanical model. The experiments prove the lower error ($< \text{0.4}\,N$) of the designed regression by comparing it with other approaches, such as artificial neural network and long short-term memory. It should be also mentioned that the antinoise ability is an important index to apply this technique in practical applications. Hence, we also add different Gaussian noises into the dataset to demonstrate the robustness against measurement noises by using the proposed model. Finally, it demonstrates the performance of the proposed algorithm using the Myo controller and KUKA LWR4+ robot.