Task-space separation principle : from human postural synergies to bio-inspired motion planning for redundant manipulators
Date of Issue2017-05-05
School of Mechanical and Aerospace Engineering
Robotics Research Centre
The apparent conflict between posture and movement, especially in the presence of redundant degrees-of-freedom(DOF), resulted in mutually-exclusive theories and models of human motor control and, to date, a unifying picture of how the brain manages to control both posture and movement is still lacking. In presence of kinematic redundancy, i.e. whenever multiple postures are available to satisfy a given task constraint, numerous experimental studies highlighted the existence of postural synergies: on average humans adopt a unique posture for a given task constraint. Several computational models have shown that postural synergies can be predicted via (local) constrained optimization of posture-dependent cost functions. However, often, these models are static and unable to predict movement generation. Differently, computational models capturing human-like movement features, such as straight-line hand paths and bell-shaped velocity profiles, have been traditionally formulated according to the optimal control framework. As such, these models usually lead to path-dependent terminal postures (i.e. at the end of the movement) and therefore are unable to capture postural synergies. This thesis proposes the Task-space Separation Principle and a general computational framework for posture and movement planning for redundant manipulators. The problem of kinematic redundancy is framed as a constrained optimization problem, traditionally solved in robotics via the method of Lagrange multipliers (LM). It is shown that LM act as task-space force fields that in general can be separated into a static (configuration-dependent) component responsible for postural control and a dynamic (velocity-dependent component) responsible for movement planning, leading to a novel extension of the Separation Principle previously proposed in human motor control literature. In particular, by generalizing the dynamic force field to any task-space force field policy, it is shown that the proposed approach generalizes and extends several computational models proposed in robotics as well as in neuroscience. The proposed framework is applied to the (redundant) task of pointing with the human wrist and it is shown that it can capture the experimental motor strategies (i.e. both posture and movement features) displayed by human subjects.