Model-Predictive Control of a flexible spine robot

Abstract

The Underactuated Lightweight Tensegrity Robotic Assistive Spine (ULTRA Spine) project is an ongoing effort to develop a flexible, actuated backbone for quadruped robots. In this work, model-predictive control is used to track a trajectory in the robot's state space, in simulation. This is the first work that tracks an arbitrary trajectory, in closed-loop, in the state space of a spine-like tensegrity robot. The state trajectory used here corresponds to a bending motion of the spine, with translations and rotations of the three moving vertebrae. The controller uses a linearized model of the system dynamics, computed at each timestep, and has both constraints and weighted penalties to reduce linearization errors. For this robot, which measures 26cm × 26cm × 45cm, the tracking errors converge to less than 0.5cm even with disturbances, indicating that the controller is stable and could be used on a physical robot in future work.