| Popis: |
When robots are integrated into the real world, chances are they will not be able to completely avoid situations in which they are bumped or pushed unexpectedly. In these situations, the robot could potentially damage itself, damage its surroundings, or fail to perform its tasking unless it is able to take active countermeasures to prevent or recover from falling. One such countermeasure, referred to as reactive stepping, involves a robot taking a series of steps in order to regain balance and recover from a push. Research into reactive stepping typically focuses on choosing which step to take.This thesis proposes a machine learning approach to reactive stepping. This approach leverages neural networks to calculate a series of steps that return the robot to a stable position. It was theorized that the robot would become stable if it always chose the step resulting in the highest reduction of energy. Theories were tested using a compass model that incorporated parameters and constraints realistic of an actual humanoid robot. The machine learning approach using neural networks performed favorably in both computation time and push recovery effectiveness when compared with the linear least squares, nearest interpolation, and linear interpolation methods. Results showed that when using neural networks to calculate the best step for an arbitrary push within the defined range, the compass model was able to successfully recover from 97% of the pushes applied. The procedure was kept very general and could be used to implement reactive stepping on physical robots, or other robot models. |
