Robotics

Autonomous Pollination

In my thesis research on robotic pollination, I utilized my collective robotic experience to develop a machine-learning methodology for the precision pollination of flowers regardless of flower orientation. This model utilizes YOLOv8 and a fine-tuned CNN/MLP to produce end-effector velocities based on a depth image and joint information alone.

Data Collection

To train the model, I developed a custom data collection autonomy by recording manipulator trajectories in reverse. The autonomy collects data by allowing the manipulator to start at some user-defined position and traverse to and from sampled poses, collecting training data along the way. To deploy the autonomy ROS2, C++, Python, MOVEIT2, Docker, and Gazebo were utilized.

Model

The machine model utilized YOLOv8 to produce bounding boxes that are then passed to a CNN to generate controls. To integrate object positional information, manipulator-to-flower displacements are generated and used. The model's output creates output end-effector velocities and orientation, which controls the manipulator joints utilizing its Jacobian.

Results

To test the model, a remote was utilized in place of a flower, and the UR5 manipulator from WVU IRL was used. 50 trials were conducted and a success rate of 88% was demonstrated.

DataCollection.mp4

Data Collection

Success1_full.mp4

Visual Servoing

Tracking_full.mp4

Tracking

Air Hockey Table Manipulation

In my first year of graduate school, I was awarded a NASA space grant to analyze imitation learning methods for mobile robotic manipulation. The environment for these tests was based on the physics of an Air Hockey table. Three different behavioral cloning based methods were analyzed within the simulation and basic behavioral cloning was deployed on a real robot.

Simulation

trial_1.mp4

Behavioral Cloning

trial_2.mp4

DAgger

trial_3.mp4

Image Behavior Cloning

Real World Results

Air Hockey Real.mp4

Loopy

Loopy is a robotic platform that's inspired by the biological process of morphogenesis, allowing it to generate its own body. My involvement with this project was for the experimental design and integration. Results were demonstrated and published at IROS 2023.

Model

The controls are based on Turing patterns which allow the morphogens (chemicals) to react and diffuse throughout individual cells (servos). This allows Loopy to generate its body shape.

Page updated

Google Sites

Report abuse