Legged swimming robots, such as the Aqua platform, present a difficult control problem due to fluid dynamics,
force effects of varying time-scales, and the general challenge of working underwater. As a postdoc at McGill's
Mobile Robotics Lab, I have lead a project to improve Aqua's control systems, which has involved the use of
state-of-the-art techniques in learning from reinforcement.
We began by extending Aqua's control system to better handle 3D motions
such as barrel rolls and cork-screws. We have also considered the HRI problem facing a diver programming the robot to execute such
motions productively. Our results appeared at IROS 2014 ([3D Trajectory Synthesis and Control for a Legged Swimming Robot]).
Recently, we have demonstrated the learning of novel gait types from experience, based on Gaussian Process dynamics models
learned from experience data. This work was a best paper award finalist at ICRA 2015 ([Learning Legged Swimming Gaits from Experience]).