Integrated affordance detection and humanoid locomotion planning

Abstract

A humanoid robot navigating in an unstructured environment requires knowledge of the affordances which allow it to make contact with the environment. This knowledge often comes from a perception system, which processes data from 3D sensors such as LIDAR and extracts available areas for the robot to make contact. Because perception systems run independently of the robot's planner, without knowledge of the robot's goal, they must process the entire visible area. In large environments, or those with complex geometry, the perception system may spend significant time processing areas of the environment that the planner will never consider visiting. By integrating the perception process with the planner, we are able to improve the speed with which the robot can compute a motion plan by only processing those areas of the environment which are considered by the planner for navigation. Two experiments with simulated and real-world point cloud data suggest that our framework can produce comparable plans up to seven times faster than the perceive-then-plan approach.