I recently implemented some of the robust neural network architectures developed in our lab at ACFR on the Qube Servo. The GIF on the left shows the Qube being controlled by a Youla-REN policy, which I trained using a custom deep RL pipeline based on the MJX physics simulator.

Youla-REN policies augment an existing stabilising controller with a recurrent equilibrium network (REN) to improve performance while preserving stability. I started by designing a classical swing-up controller, combining an energy-pumping controller with an LQR, and added an extended Kalman filter for state estimation. Putting it all together with RENs, deep RL, and domain randomisation results in the swing-up policy in the GIF.

It's not every day that one has a bipedal robot on hand for a passion project. Luckily, we have a Cassie robot at the ACFR which a number of our PhD student and postdocs use for their research in in motion planning and control.

I joined the Cassie team at the start of my PhD to assist with designing, testing, and implementing locomotion and balancing policies. Along the way, I learned a lot about interfacing, robotic software engineering (ROS), and state estimation with discrete contact events. Most importantly, I had the opportunity to get hands-on with a range of interesting control architectures based on inverse kinematics and task-space inverse dynamics control.

I also got to embark on a very unnecessary but entirely enjoyable side-project to control Cassie directly from Julia via its native C/C++ interface. Very unnecessary... but a great deep-dive into low-level interfacing between programming languages, and it worked! (GIF on the left.)

During my Bachelor's thesis, I had the opportunity to design and build a vibration control test-bed for Nearmap, an aerial imaging company based in Sydney. Over the course of six months, I: 1) built a piezoelectric cantilever beam system; 2) modelled the system analytically, with FEA, and via system identification; 3) designed active damping controllers with positive position feedback and LQG control; and 4) implemented the controllers in real-time on an STM32 microcontroller.

One of my earliest research positions was as part of a team of students working to a develop a commercial off-the-shelf (COTS) star tracker for attitude determination on small satellites. At the time, there were very few commercial star trackers that were small yet accurate enough for CubeSats. Our goal was to develop our own, in-house instrument for deployment on University of Sydney satellites.

I worked on preliminary designs of the CROSS star tracker. CROSS has since been flown to space and deployed on board the CUAVA-2 satellite.