Kinodynamically Consistent Motion Retargeting for Humanoids

Abstract

Human-to-humanoid motion retargeting is an important tool to generate human-like humanoid motions. This retargeting problem is often formulated as a Cartesian control problem for the humanoid from a set of task points in the captured human data. Classically, Cartesian control has been developed for redundant systems. While redundancy fundamentally adds new sub-task capabilities, the degree to which secondary objectives can be faithfully executed cannot be determined in advance. In fact, a robot that exhibits redundancy with respect to an operational task may have insufficient degrees of freedom (DOFs) to satisfy more critical constraints. In this paper, we present a Cartesian space resolved acceleration control framework to handle execution of operational tasks and constraints for redundant and nonredundant task specifications. The approach is well suited for online control of humanoid robots from captured human motion data expressed by Cartesian variables. The current formulation enforces kinematic constraints such as joint limits, self-collisions, and foot constraints and incorporates a dynamically-consistent redundancy resolution approach to minimize costly joint motions. The efficacy of the proposed algorithm is demonstrated by simulated and real-time experiments of human motion replication on a Honda humanoid robot model. The algorithm closely tracks all input motions while smoothly and automatically transitioning between regimes where different constraints are binding.