Hayley Cashdollar

Summary:
For my Advanced Robotics class, I am collaborating with a CS robotics lab to understand how the selection of priors (quantity, type, proximity) influences the outcome of one-shot learning.

Abstract:
Given a linearized system with a quadratic cost function, it is possible to determine the system parameters given prior information about the state, action, and next state. The priors can be fit to a Gaussian model using Baysian statistics. In many contact-rich applications… Show more

Summary:
For my Advanced Robotics class, I am collaborating with a CS robotics lab to understand how the selection of priors (quantity, type, proximity) influences the outcome of one-shot learning.

Abstract:
Given a linearized system with a quadratic cost function, it is possible to determine the system parameters given prior information about the state, action, and next state. The priors can be fit to a Gaussian model using Baysian statistics. In many contact-rich applications such as robotic manipulation, a Gaussian model is not sufficient to represent the boundaries between different linear modes. In this case, it is preferable to use a Gaussian Mixture Model (GMM) that combines multiple Gaussians. In a GMM, each Gaussian cluster represents a single linear mode. This project explores how the characteristics of the clusters in a GMM such as quantity of data and proximity to other clusters influence the outcome of Guided Policy Search (GPS).

Skills:
- Develop MATLAB simulation of robotic manipulation (built on top of an existing simulation of the robot)
- Theoretical understanding of LQR, iLQR, GMM, Baysian statistics, Guided Policy Search
- Write MATLAB scripts for evaluating the success of different sets of priors Show less

Summary:
As a part of my MS work, I tuned the control system for a robotic exoskeleton. I also redefined the Finite State Machine to allow for a new state called “Powered Double Stance.” The addition of this state reduced the need for shoulder strength in the exoskeleton pilot - instead of a pilot propelling herself forward using crutches between each step, the exoskeleton provides the power to shift the pilot’s weight forward. Finally, I developed an algorithm that takes data from the… Show more

Summary:
As a part of my MS work, I tuned the control system for a robotic exoskeleton. I also redefined the Finite State Machine to allow for a new state called “Powered Double Stance.” The addition of this state reduced the need for shoulder strength in the exoskeleton pilot - instead of a pilot propelling herself forward using crutches between each step, the exoskeleton provides the power to shift the pilot’s weight forward. Finally, I developed an algorithm that takes data from the onboard hip encoders and torso IMU to determine if the exoskeleton pilot is ready to take a step forward or not. These three milestones allow for safe, continuous, forward walking without conscious input from the pilot.

Work detail:
- Utilized mBed microcontroller, programs written in C++
- Development of drivers for motors at the hip and knee
- Development of IMU driver, specification and testing of new IMU
- Live, manual tuning/analysis of a motor controller
- Re-configuring and writing a new FSM
- Electronics hardware debugging (understanding of circuits, ability to read circuit diagrams, electronics component understanding of functionality)
- Algorithm development Show less

This project involved balancing an inverted pendulum by manipulating the precession of a spinning flywheel at the top of the pendulum. This was with a team of 2 other people over the course of a semester.

Work detail:
- Programmed using LabView with a myRIO
- Design and manufacture of an inverted pendulum and two CMGs
- IMU driver design and implementation
- IMU filtering techniques for noise and position
- Application of ZPET control
- Servo sensor re-wiring for… Show more

This project involved balancing an inverted pendulum by manipulating the precession of a spinning flywheel at the top of the pendulum. This was with a team of 2 other people over the course of a semester.

Work detail:
- Programmed using LabView with a myRIO
- Design and manufacture of an inverted pendulum and two CMGs
- IMU driver design and implementation
- IMU filtering techniques for noise and position
- Application of ZPET control
- Servo sensor re-wiring for external position feedback Show less