E-Thesis 391 views 512 downloads
Learning Algorithm Design for Human-Robot Skill Transfer / Chunxu Li
Swansea University Author: Chunxu Li
PDF | E-Thesis – open access
A selection of third party content is redacted or is partially redacted from this thesis due to copyright restrictions.Download (9.55MB)
DOI (Published version): 10.23889/suthesis.54879
In this research, we develop an intelligent learning scheme for performing human-robot skills transfer. Techniques adopted in the scheme include the Dynamic Movement Prim- itive (DMP) method with Dynamic Time Warping (DTW), Gaussian Mixture Model (G- MM) with Gaussian Mixture Regression (GMR) and th...
No Tags, Be the first to tag this record!
In this research, we develop an intelligent learning scheme for performing human-robot skills transfer. Techniques adopted in the scheme include the Dynamic Movement Prim- itive (DMP) method with Dynamic Time Warping (DTW), Gaussian Mixture Model (G- MM) with Gaussian Mixture Regression (GMR) and the Radical Basis Function Neural Networks (RBFNNs). A series of experiments are conducted on a Baxter robot, a NAO robot and a KUKA iiwa robot to verify the effectiveness of the proposed design.During the design of the intelligent learning scheme, an online tracking system is de- veloped to control the arm and head movement of the NAO robot using a Kinect sensor. The NAO robot is a humanoid robot with 5 degrees of freedom (DOF) for each arm. The joint motions of the operator’s head and arm are captured by a Kinect V2 sensor, and this information is then transferred into the workspace via the forward and inverse kinematics. In addition, to improve the tracking performance, a Kalman filter is further employed to fuse motion signals from the operator sensed by the Kinect V2 sensor and a pair of MYO armbands, so as to teleoperate the Baxter robot. In this regard, a new strategy is developed using the vector approach to accomplish a specific motion capture task. For instance, the arm motion of the operator is captured by a Kinect sensor and programmed through a processing software. Two MYO armbands with embedded inertial measurement units are worn by the operator to aid the robots in detecting and replicating the operator’s arm movements. For this purpose, the armbands help to recognize and calculate the precise velocity of motion of the operator’s arm. Additionally, a neural network based adaptive controller is designed and implemented on the Baxter robot to illustrate the validation forthe teleoperation of the Baxter robot.Subsequently, an enhanced teaching interface has been developed for the robot using DMP and GMR. Motion signals are collected from a human demonstrator via the Kinect v2 sensor, and the data is sent to a remote PC for teleoperating the Baxter robot. At this stage, the DMP is utilized to model and generalize the movements. In order to learn from multiple demonstrations, DTW is used for the preprocessing of the data recorded on the robot platform, and GMM is employed for the evaluation of DMP to generate multiple patterns after the completion of the teaching process. Next, we apply the GMR algorithm to generate a synthesized trajectory to minimize position errors in the three dimensional (3D) space. This approach has been tested by performing tasks on a KUKA iiwa and a Baxter robot, respectively.Finally, an optimized DMP is added to the teaching interface. A character recombination technology based on DMP segmentation that uses verbal command has also been developed and incorporated in a Baxter robot platform. To imitate the recorded motion signals produced by the demonstrator, the operator trains the Baxter robot by physically guiding it to complete the given task. This is repeated five times, and the generated training data set is utilized via the playback system. Subsequently, the DTW is employed to preprocess the experimental data. For modelling and overall movement control, DMP is chosen. The GMM is used to generate multiple patterns after implementing the teaching process. Next, we employ the GMR algorithm to reduce position errors in the 3D space after a synthesized trajectory has been generated. The Baxter robot, remotely controlled by the user datagram protocol (UDP) in a PC, records and reproduces every trajectory. Additionally, Dragon Natural Speaking software is adopted to transcribe the voice data. This proposed approach has been verified by enabling the Baxter robot to perform a writing task of drawing robot has been taught to write only one character.