Graduated automation for humanoid manipulation

Abstract

The challenges posed by humanoid robotic manipulation in complex, remote environments have shown that fully-autonomous and fully-teleoperated approaches are impractical. In real-world environments, one encounters rapidly changing situations, localization drift, and other errors that break autonomy. Pure teleoperation is impractical due to the sheer number of degrees-of-freedom, real-time constraints such as balancing, and potentially low-bandwidth or unreliable communications between operator and platform. To address these issues, the operator must be able to seamlessly interact with the platform across multiple levels of abstraction. In this paper we present a manipulation framework underpinned by a novel system for applying graduated levels of automation to humanoid control. The level of automation used by the operator is dictated both by their preference as well as the system's own confidence in succeeding autonomously. We present experimental results using the Boston Dynamics Atlas humanoid platform.