Title: RoboSuction: Safe and Efficient Suction under Partial Observations
Abstract: Safe autonomous suction is a desirable capability for various medical applications, such as cleaning up excess blood during open surgery. Different safety constraints are present and must be respected, such as not touching the underlying surface. Moreover, planning can be challenging due to liquid dynamics and occlusions, making it difficult to estimate the unknown surface. Some prior over the type of surfaces can often be assumed, but inference of the exact surface is needed for safe and efficient cleaning of the remaining liquid.
We propose a hybrid planning approach, by using reeb graphs and gaussian processes to represent the fluid phenomena under uncertainty conditions. By exploiting this reduced fluid dynamics representation, we propose a formulation that can reduce the planning problem to a set of standard traveling salesman problems (TSP). We demonstrate that this outperforms reactive approaches for performing suctions, as well as myopic planning strategies taking the surface estimation into account.
Abstract: I will talk about recent work on modelling fracture dynamics using material-points for visual effects. I will also talk briefly about a project on enhancing facial animation with physically based muscle fibre contractions using finite elements.