Research On Several Problems Of Humanoid Robot Motion Control And Planning

The Robot merges multiple disciplines such as mechanics, electronics, computer sicence, material science, artificial intelligence and is an important indicator of a country's comprehensive potential of science and technology. In late years, robot develops very fast in the world which is taken to ignite the fire of the third industrial revolution. As one the frontier subjects, humanoid robot is also evolving vigorously. For the similarity of humanoid to a human being in many aspects like appearance, behavior, vision function, e.t.c., a humanoid is an ideal object to live with human and help them perform daily tasks. Similar to other types of mobile robots, locomotion is one of the core functions of humanoid as the basis for more complex functions. Contrast to other types of robots, a humanoid is equipped without a stationary base, possesses relative smaller support area, includes both discrete and continuous motion and is more difficult to control.This thesis takes deep research into the following important problems which affecting robot walking:1. Hydraulic actuator controlFor the variation of load during robot walking, model uncertainty as well as the nonlinear variation of system state, the conventional fixed point control like PD (Proportional Derivative) control or LQR (Linear-Quadratic Regulator) control cannot reach a satisfactory performance. Therfore, this paper proposes a Neural Network based adaptive control method. The sequential Extended Kalman Filter is used to estimate an augmented state/space vector on line and calculate the sensitive coefficients depending on the system state and unknown parameters. Optimal controllers are designed based on the sensitive coefficients. The RBF (Radial Basis Function) neural network is used to learn the mapping from sensitive coefficients to controller gains. The fully trained RBF Neural Network can predict the controller gains in whole state space. The proposed method can effectively tackle the problem of model uncertainty and nonlinearity.2. Improved inverse kinematicsThe conventional numerical iteration method of inverse kinematics has the singularity problem, i.e., for some posture configuration the Jacobian matrix cannot be inversed. Around singular configuration, although the Jacobian matrix is invertible, the iteration speed and accuracy is degenerated for the large condition number. Further, numerical iteration method adopts the previous solution as the initial condition of iteration at current time and tends to cause accumulation of errors. As another kind of numerical method, Neural Network is not subjected to such problems as singularity or error accumulation, but the predict accuracy is lower. Therefore, this thesis combines the Neural Network and numerical iteration and takes each as dominant according to the singularity condition. This method takes the advantages and avoids the disadvantages of the Neural Network and numerical iteration methods:for non-singular configuration, the Neural Network provides initial condition for the numerical iteration which optimizes the Neural Network prediction further; for singular configuration, the use of Neural Network can guarantee the convergence of the solution.3. Walk planning on complex terrainsTo walk out of laboratory into the environment of human and perform serving tasks, the humanoid must be able to walk on complex terrains or in natural environment. Currently, several problems for walk planning exist:on one side, the widely-used ZMP (Zero Moment Point) stability indicator can only be used on level ground; on the other side, the Cart-Table ZMP model constrains the CM (Center of Mass) in vertical dimension. Therefore, this thesis proposes EZMP (Extended Zero Moment Point) generalizing ZMP to uneven terrain. Through analyzing the relation between EZMP and ZMP, it is found the EZMP model is identical to ZMP if the centroidal angular momentum rate is neglected. By implementing an ordinary differential equation as constraint on the vertical motion, the ZMP kinematic model is decoupled into three independent linear differential equations. The decoupled equations incorporates Cart-Table model and extends the CM motion from planar to three dimensional space.4. Human-like walk planningTo live and work with human, it is not enough to only walk stable. The social and psychological factors should also be considered. This means robots motion should'seems' like human. The conventional human-mimic motion are through motion capture method, which does not consider the physical or dynamical imitation of human motion and tends to cause the physical infeasibility of human-captured motion applying on a robot. Foot CoP (Center of Pressure), as the only controllable external force in normal walking, is important to human/robot dynamics. Therefore, this thesis proposes a method merging the human-mimic features in limb motion and ZMP (equivalent to CoP) to generate physically feasible human-like walking. The postures of the limbs are captured using IMU (Inertia Measurement Unit) and ZMP measured using human foot-bottom EMG (Electromyogrphy). The property, advance of the EMG to ZMP is important for improving the real-time capability and walking stability of robot. For the complexity of bio-signal, the EMG signal is calibrated using force plate beforehand and online measured EMG is smoothed in real time. Based on the ZMP prediction, CM coordinate is generated using preview control and merged with the measured limb positions and postures using optimization method.5. Energy efficient walk planningNowadays, the most advanced (fully actuated) humanoid robot can only achieve 1/20 to 1/10 of human energy efficiency at most. The energy consumption of robot includes actuator friction, the electric heat loss and mechanical energy dissipation in walking motion. One possible solution for improving the energy efficiency is developing actuators with lower friction and heat dissipation. However, this solution is infeasible until new materials or relating technologies are fully developed. Another solution for saving energy is planning appropriate walking motion which can take more advantage of mechanical energy to improve the whole energy efficiency. The intuitive method is to make robot mimic human natural walking. But this method is hard to use for the dramatic difference of structure and mass distribution between human and robot. As an alternative, passive walk can overcome those problems with high mechanical energy efficiency. Therefore, this thesis studies on the effects of mimicing passive walk on the energy efficiency. First, to reduce knee torque, an optimal controller is designed to control the vertical CM motion, which is combined with the preview controller in horizontal directions to realize knee-stretched walk. Then, based on the passive dynamic model of robot, passive walking gait is generated using Virtual Gravity principle. The two kind of walking gaits are merged by different percentage. Finally, consideing the effects of actuator friction and other factors on the energy efficiency, Reinforcement Learning is used to optimize the energy consumption of the mixed gait. To guarantee the walking stability, a stability optimization factor is added in the cost function. By observing the effects of mixing gaits at different percentage and online learning on energy efficiency, useful conclusions are reached, which can be referred to improve the robot battery endurance.The research on several imprtant problems of humanoid field in this thesis would contribute to improving the control performance and whole body coordination of humanoid in taking complex motions, enhancing motion naturalness and reducing energy consumption. The methods in this thesis are of value in both theoretical and practical aspects.