Every motion made by a moving object is either planned implicitly, e.g., human natural movement from one point to another, or explicitly, e.g., pre-planned information about where a robot should move in a room to effectively avoid colliding with obstacles. Motion planning is a well-studied concept in robotics and it involves moving an object from a start to goal configuration. Motion planning arises in many application domains such as robotics, computer animation (digital actors), intelligent CAD (virtual prototyping and training) and even computational biology (protein folding and drug design). Interestingly, a single class of planners, sampling-based planners have proven effective in all these domains.
Probabilistic Roadmap Methods (PRMs) are one type of sampling-based planners that sample robot configurations (nodes) and connect them via viable local paths (edges) to form a roadmap containing representative feasible trajectories. The roadmap is then queried to find solution paths between start and goal configurations. Different PRM strategies perform differently given different input parameters, e.g., workspace environments and robot definitions.
Motion planning, however, is computationally hard - it requires geometric path planning which has been shown to be PSPACE hard, complex representational issues for robots with known physical, geometric and temporal constraints, and challenging mapping/representing requirements for the workspace environment. Many important environments, e.g., houses, factories and airports, are heterogeneous, i.e., contain free, cluttered and narrow spaces. Heterogeneous environments, however, introduce a new set of problems for motion planning and PRM strategies because there is no ideal method suitable for all regions in the environment.
In this work we introduce a technique that can adapt and apply PRM methods suitable for local regions in an environment. The basic strategy is to first identify a local region of the environment suitable for the current action based on identified neighbors. Next, based on past performance of methods in this region, adapt and pick a method to use at this time. This selection and adaptation is done by applying machine learning.
By performing the local region creation in this dynamic fashion, we remove the need to explicitly partition the environment as was done in previous methods and which is difficult to do, slows down performance and includes the difficult process of determining what strategy to use even after making an explicit partitioning. Our method handles and removes these overheads.
We show benefits of this approach in both planning robot motions and in protein folding simulations. We perform experiments on robots in simulation with different degrees of freedom and varying levels of heterogeneity in the environment and show an improvement in performance when our local learning method is applied. Protein folding simulations were performed on 23 proteins and we note an improvement in the quality of pathways produced with comparable performance in terms of time needed to build the roadmap.
- Amato, Nancy Professor - Term Appointment