Dept. of Scientific Computing
Florida State University
"Robots with Legs and Robots without! Fun Applications of Intelligence and Planning"


Imagine trying to run across a field with trees and bushes dotting the ground. How do you determine which obstacles to go around (or through), how do you pick which path is best, is one path better than another at high speeds? Now imagine that you actually don't know how exactly your legs work. Can you safely navigate your original path, is it even feasible?

Oh, and assume that you have bad eyesight.

Motion planning for robots on unstructured terrains is oftentimes very difficult. Field performance often suffers from limited perception, unknown terrain characteristics and hardware constraints; all of which seriously impacts the robot’s behavior. A combination of good control techniques and accurate modelling can mitigate much of the impact of disturbances to the robot.

This talk will be about using neural networks as robot motion models, A*-type algorithms to search space and model predictive control to robustly correct perturbations. The techniques discussed are fast enough for use on physical hardware, and they have the advantages of taking into account complex models of the robot’s motion to generate a direct trajectory. This method is light-weight and can be combined with higher-level task planners to solve many types of planning problems.