Intelligent Quadruped: Simplified RGBD-Based Autonomous Navigation for a Quadruped Robot

Abstract

To move out of supervised research labs and be useful to mainstream users in day-to-day life, robotic systems need to be made more affordable and more reliable. This thesis addresses the problem of developing autonomous navigation systems that enable safe traversal through cluttered terrains at low computational cost. We design a lightweight computer vision platform for the Minitaur, a GhostRobotics quadruped robot, consisting of an RGB-D camera and a computational unit. This platform allows the implementation of techniques that can identify and react to obstacles in real time. To compensate for the sparse depth data from the camera, our approach approximates the ground truth of the full depth matrix by interpolating with either radial basis functions or Voronoi polygons. Furthermore, we apply an adaptive grid sizing algorithm to the interpolation results to reconstruct a contrast-rich depth image. The reconstructed depth data is used by a vectorized 'Find the Gap' method to direct the Minitaur and ensure collision-free navigation. This software procedure is then tested on three different sensor mount designs with varying degrees of freedom to compare the reliability and navigation capability between designs. A total of 3000 computer simulations and 50 physical trials are performed to test our design implementations. Both simulations and physical trials provide strong evidence in support of the rotational head design, especially in more complex environments. The results indicate that RGB-D based sensor platforms for highly agile robotic systems, such as the Minitaur, have the potential to act as affordable, low complexity sensor alternatives for autonomous navigation.