Model Refinement for Terrain Responsive Planning on a Dynamic Running Quadruped

Abstract

Running legged robots present several challenges when motion planning, these challenges often stem from the difficulty in predicting motion due to the innate complexity of the mechanical system, persistent effects of terrain and foot-ground interactions. While reasonable approximations of the inherent motion models of the Ghost Robotics Minitaur platform can be learned through data-driven approaches, system mechanical robustness and the requisite experimental time discourages running a full battery of experiments to determine a unique model for each considered terrain. This paper discusses the development of turning maneuvers on the quadruped robot Minitaur and the approach taken to adapt a learned model for new terrains. The prediction model was generated through data collected in indoor experiments using a VICON motion capture system and adapted through use of a correction factor for new terrains. Resulting motion planning differences show that the addition of this factor improves the planning model and reduces the computational burden of replanning.