Motion planning, the task of determing a path for a robot from a start to a goal position while avoiding obstacles, is a requirement for almost all robot applications. For robots with many degrees-of-freedom, motion planning must be performed in complex, high-dimensional spaces. In these high-dimensional spaces, many feasible robot configurations are chained together to form a motion plan, requiring hundreds or thousands of costly collision checks.

Repeated collision checking is computationally expensive, taking up to 90% of modern motion planners’ computation time. Most motion planning researchers look into reducing the number of collision checks, but not much effort is expended into accelerating the collision checks themselves.

We investigate machine learning models that may be used as a fast proxy to standard collision checking and distance-to-collision paradigms. Our proxy collision detection algorithm, Fastron, accurately determines a robot’s collision status an order of magnitude faster than state-of-the-art collision checking methods, efficiently updates in response to a changing environment, and scales well with large numbers of collision objects. Current research investigates further speed and accuracy improvements, GPU parallelization, probability-of-collision predictions, new feature spaces, and collision-free configuration generator models.