Optimization-based whole-body control and reactive planning for a torque controlled humanoid robot

2018-07-26In this work, we explore computationally lightweight optimization-based methods for planning and control of a humanoid robot. In doing so, we focus on the important role of contact for the task of balancing and walking for a bipedal system. This thesis puts a strong focus on real robot implementations and will address questions that come with working with real complex systems. We begin by looking into whole body control techniques for torque controlled humanoids. While many research groups have converged to using Quadratic Programming (QP) based inverse dynamics techniques, we explore to what extent computationally simpler optimal control strategies work well on real systems. We show how linearizations of the full dynamics can be projected into contact consistent constraints to compute time-invariant optimal feedback gains. We then show how high level task goals can be mapped into this framework. We evaluate this computationally lightweight control policy in a variety balancing tasks and show that for significantly less computational effort, disturbance rejection competitive with the state of the art methods can be achieved on the real robot. We then extend this work to use linearizations around different contact configurations and postures to stabilize more complicated motions that involve changing contacts such as walking. We conclude by developing an online walking pattern generator that generates hand contact plans when adjusting the center of mass (CoM) and footstep location alone is not sufficient. The lightweight formulation presented allows for hand contact locations and force trajectories to be determined in real-time without predetermining the specific contact location or timing