Trajectory optimization for dynamic aerial motions of legged robots

Abstract

A novel framework for planning and executing dynamic aerial motions for legged robots is developed. These dynamic capabilities allow legged robots to conquer challenging obstacles like gaps and hurdles that cannot be traversed via standard walking and running gaits. The framework consists of two main steps. First, a motion planning step uses trajectory optimization to generate a dynamically feasible motion of the robot that achieves a desired behavior. The desired behavior, which comes from a higher-level planner or a human operator, can specify an arbitrary 3D motion task such as jumping onto a platform or performing a front flip. The trajectory optimization simultaneously optimizes the centroidal dynamics and joint-level kinematics of the robot to plan general 3D motions. Novel actuator constraints are imposed on the optimization that ensure all planned motions are feasible for implementation on hardware, and a two-stage formulation of the optimization automatically generates dynamically-informed warm starts to the optimization that dramatically reduce solve times. The second step of the framework is a unified whole-body controller that tracks these planned motions. The whole-body controller uses a prioritized task hierarchy that is optimized for robust tracking and safe landing of dynamic aerial motions. The ability of the proposed framework to reliably produce 3D aerial motions such as running jumps, barrel rolls, and flips is demonstrated on the MIT Humanoid robot in simulation and on the MIT Mini Cheetah robot both in simulation as well on hardware.