Learning Control And Generation Of Optimal Trajectory For Robot With Underactuated Joints

Different from a conventional robot's motion, there is much coupling among joints in the motion of human (or other creature). Human have talent for passing the motion energy by these coupling. And also, human have talent for controlling and passing the motion by the limit and the flexibility of joints. Compared with robot, human can move with better efficiency. In our research, we simplified the mode of such motion transfer as a dynamics coupling mode of actuated joints and unactuated joints (or named as passive joints / free-joints). In order to simulate human's motion mode, the actuated joints drive those unactuated joints by dynamics coupling. The research on control of under-actuated system is signality not only for research on bionic mechanics, but also for other field, for example, space robot, low-cost robot, including improving the performance of conventional robot at high speed. Due to the unactuated joints, the motion control of an under-actuated system is much more difficult than that of a conventional system. The unactuated joints can't be controlled directly, so it is an incomplete controllable system. Smooth feedback control is invalid for such system. According to same reason, the motion plan for an under-actuated system is much difficult. Because the unactuated joints are driven by dynamics coupling, the motion trajectory must content not only the kinematics constraints, but also constraints of dynamics coupling. Here is another important problem involved in: whether the trajectory can be realized in a real system? The generation of trajectory depends on dynamics strictly, and it is very difficult to adjusted and compensated by feedback control. If the error of dynamics model can't be compensated, maybe the trajectory generated can't be realized, even can't assure the stability. There is optimization problem in control of under-actuated system also. We want the best trajectory according to some criterion, such as minimizing the work done by motor, minimizing the maximum of torque, etc. Although there are many methods of nonlinear optimization, due to the complexity and variety of nonlinear, it is difficult to find a universal and efficient method for different types of nonlinear system, especially for a real system. Aim at the problems above, a method based on learning is proposed in this paper. The basic idea is building a learning dynamics model of system by classic dynamics and artificial neural network, and then control the system by feedforward method based on it. In this way, the problem of lack for feedback control is overcome. At the same time, based on this learning dynamics model, a method of learning-optimizing by turns is proposed to generate the motion trajectory that can be realized in a real system. With the precondition of realizability, conventional algorithm of nonlinear optimization is improved to generate the optimal trajectory for a real robot with unactuated joints. The prime principle, the method of learning control and the algorithm of nonlinear optimization are discussed in this paper. Experiment has been done on a golf swing robot to substantiate the efficiency and the realizability of this method. The scenario and the results of the experiment are discussed and analyzed in this paper. Under-actuated robot is a new field of robotics research. There is no ripe and systemic theory about this field. In this paper, a new method is proposed to deal with tracking control and motion plan of an under-actuated system. It is applicable to iterative or similar action of a robot. The results of experiment show that it is more efficient and practicable to a real system. Solving the key problem of motion control, the unactuated joint can be applied much more.