GIIM: Model-driven Lightweight Neural Network for Robot Force Control in Dynamic Complex Environment
Robot is known as "the pearl at the top of the crown of the manufacturing industry", and serves as an important supporting tool of intelligent manufacturing. "human-computer integration" is an increasingly important area in robotic and artificial intelligence technology: robots and humans share the working space, interact naturally, and coordinate closely to achieve human-computer integration. Force control is capable of realizing the combination of rigidity and flexibility through the force perception and control of the environment. For example, in robotic surgery, the scalpel can achieve effective suture of trauma and avoid secondary injury to patients; in medical rehabilitation tasks, robots are used to provide auxiliary movement for patients, and promote the recovery of muscle strength and endurance. Therefore, the research on force control technology can effectively improve the task execution ability of the robot.
Due to the direct interaction with environment, the force control performance of the robot is directly affected by the external environment. With the growing complexity of the environment, force control of a robot is facing great challenges: in the process of force control, the influence of dynamic obstacles and its own physical constraints on the robot should be considered. In addition, the variable payload and uncertain parameters would increase complexity .
For this reason, GIIM has proposed a model-driven neural network to improve force control performance. Given a specific task, the basic procedures of the model-driven neural network are summarized into three steps:
A model family is first constructed based on the task backgrounds (e.g., objective, physical mechanism and prior knowledge). It is a family of functions with a large set of unknown parameters, amounting to the hypothesis space in machine learning. Different from the accurate model in the model-driven approach, this model family only provides a very rough and wide definition of the solution space. It has the advantage of a model-driven approach but greatly reduces the workload of accurate modeling.
An algorithm family is then designed for solving the model family and the convergence theory of the algorithm family is established. The algorithm family refers to the algorithm with unknown parameters for minimizing the model family in the function space. The convergence theory should include the convergence rate estimation and the constraints on the parameters, which assure the convergence of the algorithm family.
The algorithm family is unfolded to a deep network with which parameter learning is performed as a learning approach. The depth of the network is determined by the convergence rate estimation of the algorithm family. The parameter space of the network is determined by the parameter constraints. All the parameters of the algorithm family are learnable. In this way, the topology of the deep network is determined by the algorithm family, and the deep network can be trained through back-propagation.
Fig. 1 Framework of the control structure of the model-driven lightweight neural network for robot force control
Taking the influence of dynamic and uncertain obstacles into consideration, 3D vision technology is used to monitor the robot’s surrounding environment. The static or dynamic obstacles introduced by humans are captured by removing the image background through background difference method. An inequality-constraint-type obstacle avoidance scheme is established based on a dynamic system theory, which is merged into a unified problem description framework. Considering the unknown parameters of the contact process, the Cartesian impedance control strategy with adaptive parameter learning in the outer loop is designed. Defining the energy consumption of the system as the optimization index, combined with the physical constraints of the joints of the system, the force control problem description with constraint optimization form is finally obtained.
Experiment results show that when the joint angle and velocity of the robot are limited, the control error of the contact force on the surface with unknown stiffness is less than 1%, and the energy consumption is saved by 16%. This method greatly simplifies the neural network structure and effectively improves the adaptability of the robot system.
This research is published in IEEE transactions on industrial electronics (DOI: 10.1109/TIE.2020.2970635).