Passive motion paradigm implementation via deep neural networks: analysis and verification

In recent years, passive motion paradigms (PMPs), derived from the equilibrium point hypothesis and impedance control, have been utilised as manipulation methods for humanoid robots and robotic manipulators. These paradigms are typically achieved by creating a kinematic chain that enables the manipulator to perform goal-directed actions without explicitly solving the inverse kinematics. This approach leverages a kinematic model constructed through the training of artificial neural networks, aligning well with principles of cybernetics and cognitive computation by enabling adaptive and flexible control. Specifically, these networks model the relationship between joint angles and end-effector positions, facilitating the computation of the Jacobian matrix. Although this method does not require an accurate robot model, traditional neural networks often suffer from drawbacks such as overfitting and inefficient training, which can compromise the accuracy of the final PMP model. In this paper, we implement the method using a deep neural network and investigate the impact of activation functions and network depth on the performance of the kinematic model. Additionally, we propose a transfer learning approach to fine-tune the pre-trained model, enabling it to be transferred to other manipulator arms with different kinematic properties. Finally, we implement and evaluate the deep neural network-based PMP on the Universal Robots, comparing it with traditional kinematic controllers and assessing its physical interaction capabilities and accuracy.