I'm Ryan J. Hayward, and I'm currently a PhD candidate at the University of St Andrews, supervised by Kasim Terzić. I focus on pure imitation learning for robotics, specifically on long-horizon tasks which can be split into sub-tasks and chained together. I'm also interested in teleoperation for robot demonstration collection, and have been developing interfaces between VR hardware such as the Meta Quest and the Vive Pro 2 and robotics control code for this exact purpose.
My research interests include hierarchical/chained imitation and reinforcement learning for robotics tasks, but I also have a keen interest in VR. I am currently working on approaches which use pure VR demonstration data to learn tasks with a focus on success rate. Although my work has been in simulation (MuJoCo, PyBullet, Unity) thus far, I will be integrating my VR control system with our UR3e robot soon for real-world applications. More coming soon.
An imitation learning method which learns each sub-task and its termination probability jointly for long-horizon sequential robotic tasks, from demonstrations with sub-task labels. TASC consistently outperforms several state-of-the-art methods on well-known Gymnasium-Robotics baselines when learning from segmented sub-tasks.
I also like to make arcade-style games with Unity and hand-program them with C#. Feel free to play them in your browser.
Use the mouse to drag a paddle to bounce a ball and destroy clusters of bricks. Make big combos for big powerups, and discover procedurally generated bonus levels. The difficulty of the game increases over time, and bonus levels are generated procedurally by combining a set of algorithmic templates with the dimensions of the brick grid (which increases in size over time).
Use the WASD keys (or arrow keys) to fly a spaceship through procedurally generating pipes. Try not to hit the walls. Pipes are rendered from scratch with vertices from the equation for a torus, connected together with procedural triangles and shaded in a wireframe pattern.