Planning and Control in Face of Uncertainty with Applications to Legged Robots

The present work aims at motion control and planning for autonomous robots, specifically legged robots. Motion control and planning have always been an important component for creating autonomous robots. There has been significant progress in this field for UAVs and wheeled robots since the early ages of robotics. However, motion control and planning for legged robots is still a challenging problem and some of the recent, notable robotic challenges have demonstrated the necessity of more robust and online approaches. To this end, our work in this dissertation revolves around three main challenges of motion control and planning in legged robots namely real-time planning, robustness of the control structure and adaptation to unknown or neglected dynamics. One of the essential requirements for robust planning in real world applications is the capability of finding solutions in a real-time fashion to adjust the plan with the current state measurements. Many of today’s online approaches have achieved this efficacy through task decomposition and model reduction approaches, since whole-body approaches are often significantly slower to be run online. To address this problem, we have proposed an efficient algorithm for whole-body planning of floating based robots with switching dynamics. Our proposed method is based on an optimal control approach which synthesizes an optimal feedback control policy for continuous inputs and optimizes the switching times in between two consecutive stance modes. Through this optimization approach, we have demonstrated that a wide spectrum of motions can be generated for a quadrupedal robot with predetermined gait sequences. Furthermore, we have employed this algorithm for planning highly dynamic gaits (e.g. trotting) in a nonlinear Model Predictive Control (MPC) fashion on hardware. This is one of the earliest works on whole-body MPC for periodic gait generation of legged systems.