Conference-style papers written for graduate course research projects.
SYDE 422: Machine Learning
Learning Thought-Based Motor Control Using Gaussian Processes
Over the recent decades, brain-computer interfaces (BCI) have become a highly active area of research. Due to the nondeterministic nature of brain wave patterns, machine learning approaches are commonly used to develop intelligent systems which are designed to act on a users given thoughts. While a wide range of techniques have been used over the past few decades, support vector machines (SVM) approaches have yielded the most success in this field. Another kernel methods based machine learning approach similar to SVM can potentially improve the classification accuracy of the machine learning approach. The output of a GP-based algorithm is a probabilistic distribution which can be combined with external sensory information to improve the overall performance of the BCI system. The results of a three class multivariate system was used to demonstrate the efficacy of GP-based machine learning approaches. While GP was capable of achieving similar accuracies as SVM during the training phase, the performance was slightly lower during the validation phase. The performance trade off can be attributed to poorly tuned hyperparameters for the GP-based algorithm. However, the main benefits of coupling the probabilistic output with the GP-based algorithm can be explored in future research to obtain the overall performance benefits when compared to other approaches.
ECE 687: Robot Dynamics and Control
Accurate Determination of Joint Angles from Inertial Measurement Unit Data
This paper describes an approach to accurately determine joint angles from the sensor fusion of gyroscopic and acceleration data by implementing an Extended Kalman Filter (EKF) In similar research it has been determined that the sensor fusion of inertial measurement unit (IMU) data gives rise to drift errors due to integration. The magnitude of this error increases further as you traverse down the kinematic chain. The Extended Kalman Filter adjusts for this non-linear drift error by using a state prediction model. Through simulations based on measured data, it was shown to be an effective in reducing drift errors in calculating the shoulder joint angle. Furthermore, it was also shown that the impact of drift errors further down the kinematic chain (i.e. the elbow) were also reduced by training the EKF with a dynamic model prior to running simulation.
ECE 780: Humanoid Robotics
Energy Efficient Gait Control for Minimally Actuated 3D Active Dynamic Walking
Biped locomotion is a highly active area of research within the field of humanoid robotics. A new and emerging field of highly energetic walking mechanisms known as passive dynamics have become a topic area of interest. Typically, passive mechanisms exhibit stable walking cycles on a sloped incline where gravity is doing work. However, in order to bring the benefits of passive dynamics into more realistic robots, level walk must be realized. The field of virtual passive dynamics presents a novel idea whereby the work done by gravity is emulated by minimal actuation. This provides a means for energetically efficient walk on level ground. However, the research thus far has focused on virtual passive dynamics in the 2D sagittal plane. A model is developed in this paper to extend the system into 3D by considering the lateral dynamics. It was found that controlling the gait in 3D is highly challenging and there exists many parameters which greatly influence the overall stability of the system. Ultimately, a detailed analysis on the system energy level is required to produce a stable 3D walking mechanism for level ground.
ME 780: Autonomous Mobile Robots
Bipedal Locomotion for a Lower Body Humanoid Research Platform
Bipedal locomotion is a highly active area of re- search within the field of humanoid robotics. Thus far, the most popular control strategy used to achieve walking in robots has been based on the Zero-Moment Point (ZMP) criterion. In order to achieve stable gait cycles for a humanoid robot, the walking control strategy must track a stable COM trajectory which keeps the ZMP location within the region of foot support. A simplified motion model for the walking robot is obtained by approximating the full system as a linear inverted pendulum. Through force senors mounted on each foot and an inertial measurement unit (IMU) mounted at the torso, the multi-layered control strategy is capable of computing the instantaneous ZMP location and correcting the robots motion (online) while executing stable walking. The full system is implemented in simulation to analyze the performance of the approximated motion model and robustness to additive noise in the sensor models.
ECE 682: Multivariable Control Systems
Linear State Feedback Controller Design for Bipedal Locomotion
Bipedal locomotion is a highly active area of re- search within the field of humanoid robotics. Thus far, the most popular control strategy used to achieve walking in robots has been based on the Zero-Moment Point (ZMP) criterion. In order to achieve stable gait cycles for a humanoid robot, the walking control strategy must track a stable COM trajectory which keeps the ZMP location within the region of foot support. A simplified motion model for the walking robot is obtained by approximating the complex dynamics of the system as a 3D linear inverted pendulum (3D LIPM). This approximation provides a casual, linear time-invariant (LTI) system which can be controlled using linear techniques. It was determined that the state space model of the 3D LIPM is both controllable and observable. However, the system is inherently unstable due to poles in the right half plane. In order to stabilize the system, a linear state feedback system is designed. The final control strategy is implemented in simulation on a 14DOF lower body humanoid research platform which closely models the kinematic and dynamic parameters of a physical robot. It was determined that the linear control strategy utilizing state feedback was successful in keeping the ZMP within the region of foot support.