Dynamic Primitives in Human Manipulation of Complex Objects

Abstract

Humans are remarkably adept at interacting with complex objects and display an impressive level of dexterity in manipulating underactuated and nonlinear objects. Inspired by the task of quickly transporting a cup of coffee while inducing minimum internal oscillations, this study investigated the nature of human interactions with a complex object. An open-loop control strategy consisting of composition of submovement and impedance dynamic primitives was proposed and simulated. The effect of shaping command inputs by implementing internal models with different levels of structure was analyzed.

Previous work has proposed maximum smoothness optimization-based criteria to describe observed human interactions with flexible objects: minimum crackle of object (MCO) and dynamically constrained minimum jerk of hand (DCMJH). These models fail to reproduce human subject data in two ways: experimental hand velocities take on asymmetric bimodal profiles with shorter move durations, while the optimization-based models predict purely symmetric velocity trajectories. Additionally, the local minimum velocity between peaks was observed to increase with shorter move durations, while the optimization-based models predict an increasingly negative velocity minimum. Finally, by their nature these models serve a descriptive function but do not account for how motions may be planned and executed by the human central nervous system.

Movement generation using an open-loop strategy of dynamic primitive composition based on an internal model was shown to be a competent descriptor of observed behavior and superior to the optimization-based models. The proposed models generated asymmetric bimodal velocity trajectories, and the choice of internal model influenced the relationship between minimum inter-peak velocity and movement duration, with some models reproducing the negative correlation observed in human subject data. Simulations that used internal models of a lower order than the simulated physical plant and employed feedforward force input fit observed motions better and with more biologically-feasible impedance values than those that employed more precise internal models and neglected feedforward force. These results suggest that internal models may play a key role in human interaction with complex objects, and that humans may rely on less detailed internal models to simplify interaction tasks.