Optimization based motion planning under task space constraints

Motion planning for robots with a high number of degrees of freedom is a difficult problem. Randomized algorithms have gained popularity as a method to alleviate the inherent problems related to the curse of dimensionality and avoid explicit computation of the free / occupied portions of configuration space. Randomized algorithms however have several disadvantages: no guarantee on optimality or completeness, as well as unnaturally looking and non-smooth trajectories. As an alternative, recent years have seen some progress in deterministic, optimization-based motion planners, e.g. CHOMP, STOMP.

A less explored feature of optimization-based planners is that they should lend easily to imposing task-space constraints on portions (or the complete) final trajectory. For example, if we need to plan a trajectory for carrying a glass of water, it would be desirable to impose a limit on the allowed tilting of the glass throughout the motion. In this thesis, we will investigate efficient ways of imposing such task-space constraints on the motion planner objective function. We will integrate constraints on goal regions and compare motion planning performance against single goal planning. Implementation of a planner from scratch is not expected in this work, but a considerable coding effort will be involved nonetheless. Implementation and testing on ABB YuMi will be included in a blue-sky scenario.

Contact: Todor Stoyanov