Research

Learning Heuristic A*

A* algorithms find global optimal path under the constraints that the heuristic function should be admissible. According to my previous research experiences, we can find the optimal path faster when the heuristic function is close to the real optimal cost to guide the expansions toward the goal. So, instead of designing the heuristic function, we trained the learning heuristic(an artificial neural network) with the training set of optimal cost between pairs of vertices. By adapting SMHA*, the proposed Learning Heuristic A* guarantees bounded suboptimality. With numerous simulations, LHA* find good path(cost close to optimal) while expanding less vertices(squeezed to left in the above figure). This work was presented at ICRA 2020.

Topological Exploration

When deploying a group of robots to explore and build map of the unknown or partially known environment, Frontier-based planning is the state of the art. Each robot will travel to the different frontiers. This algorithm requires pre-computation and identification of the frontiers. In this work, we suggested the topological exploration algorithm as an alternative method to deploy robots to the unknown regions by comparing the topology classes of the paths. This algorithm does not require any pre-computation or identification of the algorithm.

Separate and manipulate a set of objects

Cables are widely used to manipulate or transport an object in extreme conditions like rescue missions. Most of the literature of cable manipulation consider the problem when the number of robots is greater than or equal to number of object. Motivating oil-skimming operation, we manipulate a set of objects by a cable attached at each end to a mobile robot. The set of optimal(shortest) paths for the two robots are generated by consider the topology of the initial cable configuration and paths. The details of the algorithm are presented at RSS 2013.

Dynamic simulation of cable and floating object on water surface

We model the drag on the cable and floating objects. We also consider the frictionless contact between the cable-object and object-object. This simulation is a kind of implicit consideration of the penetration with rigid contact. So it is numerically stable with larger time-step and does not need to tune contact parameters like spring or damping constants. The formulation and performance of this simulator was presented at ASME IDETC 2013, Portland, OR.

Optimal trajectory generation under homology class constraints

As the homology class of the trajectory does not change with small deformation of the trajectory, we cannot find the optimal trajectory under homology class constraints by solving optimal control problem with numerical solver. However, we can formulate the optimal trajectory generation problem as a Mixed-Integer Quadratic Program. And we define the homology class of the trajectory from its coarse representation(Integer variables) to find the global optimal trajectory under homology class constraints. As the trajectory generation problem remains as MIQP, we can guarantee the global optimality and achieve anytime solution. This work was presented as CDC 2012.