Weiwei Yu

Robust Gain Scheduling Control of Air-breathing Hypersonic Vehicle via Linear Parameter Varying Technique

The design of a robust gain scheduling controller which is scheduled on Mach number and altitude for the air-breathing hypersonic vehicle (AHV) is presented. In order to capture the nonlinear airspeed and altitude dependence of AHV, a linear parameter varying (LPV) model is constructed from a set of linearization of the AHV longitudinal dynamics. The linear fractional transformation (LFT) representation of the LPV model is obtained using the graph approach so that it can be used for robust control design technique. The control synthesis structure including this LFT representation and the uncertainty model of AHV is successfully used for controller design based on D-K iteration. Simulation result of the designed controller presents excellent tracking performance independent of the fight conditions of AHV.

Gait Pattern Based on CMAC Neural Network for Robotic Applications

The main goal of this paper is to provide a general methodology and a practical approach for the design of gait pattern for biped robotic applications directly usable by researchers and engineers. This approach, which is based on CMAC neural network, is an alternative way in comparison to the traditional Central Pattern Generator. In the proposed method, the CMAC neural networks are used to learn basic motions (e.g. reference gait) and a Fuzzy Inference System allows to merge these reference motions in order to built more complex gaits. The results of our biped robotic applications show how to design a self-adaptive gait pattern according to average velocity and external perturbations.

Self-optimizing for the Structure of CMAC neural network

CMAC neural network has been widely applied on the real-time control of the nonlinear systems, such as robot control, aerocraft control and etc. However, the required memory size increases exponentially with the input dimension of CMAC, it may conduct to serious computational challenges in its on-line application. In this paper, experimental protocol is used for illustrating how the structure of CMAC influence the approximation qualities and required memory size. It is found that an optimal structure carrying the minimum modeling error could be achieved. The self-optimizing algorithm is then developed to adjust the structure of CMAC neural network in order to accomplish the minimum modeling error with minimum required memory size, without increase the structure complexness of the network.

Estimate the Kinematics with EMG Signal Using Fuzzy Wavelet Neural Network for Biomechanical Leg Application

Several linear and nonlinear models were proposed to predict the forward relationship between EMG signals and kinematics for biomechanical limbs, which is meaningful for EMG-based control. Although using nonlinear model to predict the kinematics is able to represent rational complex relationship between EMG signals and desired outputs, there exists high risk for overfitting models to training data and calculating burden because of the multi-channel variation EMG signals. Inspired by the hypothesis that CNS modulates muscle synergies to simplify the motor control and learning of coordinating variation of redundant joints, this paper proposed to extract the synergies to reduce the dimension of EMG-based control. Furthermore, the fuzzy wavelet neural network was developed to generate velocity–adapted gait by the reference gaits only with the limited set of experimental trials. The experimental results show the efficiency and robust of this approach.

Nonholonomic mobile system control by combining EEG-based BCI with ANFIS

Motor imagery EEG-based BCI has advantages in the assistance of human control of peripheral devices, such as the mobile robot or wheelchair, because the subject is not exposed to any stimulation and suffers no risk of fatigue. However, the intensive training necessary to recognize the numerous classes of data makes it hard to control these nonholonomic mobile systems accurately and effectively. This paper proposes a new approach which combines motor imagery EEG with the Adaptive Neural Fuzzy Inference System. This approach fuses the intelligence of humans based on motor imagery EEG with the precise capabilities of a mobile system based on ANFIS. This approach realizes a multi-level control, which makes the nonholonomic mobile system highly controllably without stopping or relying on sensor information. Also, because the ANFIS controller can be trained while performing the control task, control accuracy and efficiency is increased for the user. Experimental results of the nonholonomic mobile robot verify the effectiveness of this approach.

Generating Human-Like Velocity-Adapted Jumping Gait from sEMG Signals for Bionic Leg’s Control

In the case of dynamic motion such as jumping, an important fact in sEMG (surface Electromyogram) signal based control on exoskeletons, myoelectric prostheses, and rehabilitation gait is that multichannel sEMG signals contain mass data and vary greatly with time, which makes it difficult to generate compliant gait. Inspired by the fact that muscle synergies leading to dimensionality reduction may simplify motor control and learning, this paper proposes a new approach to generate flexible gait based on muscle synergies extracted from sEMG signal. Two questions were discussed and solved, the first one concerning whether the same set of muscle synergies can explain the different phases of hopping movement with various velocities. The second one is about how to generate self-adapted gait with muscle synergies while alleviating model sensitivity to sEMG transient changes. From the experimental results, the proposed method shows good performance both in accuracy and in robustness for producing velocity-adapted vertical jumping gait. The method discussed in this paper provides a valuable reference for the sEMG-based control of bionic robot leg to generate human-like dynamic gait.

Towards Robotic Semantic Segmentation of Supporting Surfaces

Perceiving the geometry of environmental structures surrounding is a crucial prerequisite for robotic understand the indoor environments autonomously. A new framework for parsing RGB-D images aimed at supporting surfaces segmentation is proposed. First, the surface normal is extracted from depth information using PCA and normal clusters with 3D mean shift clustering. Then the main planes such as floor, wall will be detected with gravity vector estimation. Finally supporting surface and its corresponding objects are segmented using graph optimization with energy functions. The approach can offer a robotic semantic segmentation for better understanding the indoor environment. The experiment results based on Berkeley 3D Object Dataset demonstrate that our framework works well on indoor RGB-D cluttered scenes.

Model-Based Workpiece Positioning for Robotic Fixtureless Assembly Using Parallel Monocular Vision System

This paper proposed to use parallel monocular vision system that could fit different robotic grasping pattern, in order to reduce the computation burden for real-time grasping control. A novel model-based workpiece positioning approach, which can solve both 3D or 2D pose estimation problem, is proposed by using the imagery template and homography matrix. The demand of workpiece template is not 3D model, but the workpiece template image, which is much easier to obtain. Moreover, as the positioning expressions based on homography matrix between the workpiece template and images, and the two camera images, are expressed as a simple formula, the proposed approach is intuitive for algorithm development.

CMAC Structure Optimization Based on Modified Q-Learning Approach and Its Applications

Comparing with other neural network based models, CMAC has been applied successfully in many nonlinear control systems because of its computational speed and learning ability. However, for high-dimensional input CMAC in real world applications such as robot, the useable memory is finite or pre-allocated, thus we often have to make our choice between learning accuracy and memory size. This paper discusses how both the number of layer and step quantization influence the approximation quality of CMAC. By experimental enquiry, it is shown that it is possible to decrease the memory size without losing the approximation quality by selecting the adaptive structural parameters. Based on modified Q-learning approach, the CMAC structural parameters can be optimized automatically without increasing the complexity of its structure. The choice of this optimized CMAC structure can achieve a tradeoff between the learning accuracy and finite memory size. At last, this Q-learning based CMAC structure optimization approach is applied on the walk pattern generating for biped robot and workpiece orientation estimation for robot arm assembly respectively.