Elbow angle generation during activities of daily living using a submovement prediction model

Naghibi, Seyedeh Somayeh; Fallah, Ali; Maleki, Ali; Ghassemi, Farnaz

doi:10.1007/s00422-020-00834-w

Cited by 1 publication

(6 citation statements)

References 32 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…These methods rely on trajectory learning, where robots learn human skills from demonstrations and later reproduce the movement through a trained model or a function approximation. Deconvolutional neural networks (DNN) (as in [10]), multilevel convolutional neural networks (as in [8]), and artificial neural networks (ANNs) (as in [18,54]) are used for function approximation; Gaussian mixture models (GMM) (as in [14,56]), hidden Markov models (HMM), and dynamic-motion-primitive (DMP) models (as in [11]) are used as models for upper-limb exoskeleton robots in the training phase. The training is achieved using expectation-maximization (EM) algorithms (as in [14,56]) and the Levenberg-Marquart algorithm (as in [11,54]).…”

Section: Methods Applied To the Currently Available Exoskeleton Robotsmentioning

confidence: 99%

“…Learning by demonstration (LbD), also known as programming by demonstration (PbD), is sometimes employed in robotic programming for complex and nonstrict motion trajectories [52]. LbD consists of two stages (Figure 4): (i) a learning phase that first acquires behavior data and encodes them through a learning model, and (ii) a reproduction phase that uses appropriate control models to reproduce similar behavior [15,53,54]. Planning based on LbD is an alternative approach to the traditional approach to path planning and IK illustrated in the previous section since it either makes it possible to avoid motion planning in Cartesian space (as in [11]), or it helps to find a unique inverse kinematics solution (as in [18]).…”

Section: Approaches Based On Learning By Demonstration (Lbd)mentioning

confidence: 99%

“…However, this approach requires learning the target joint configuration. For upper-limb exoskeletons, the generation of reference trajectories is grounded in learning models, the most common of which are neural networks (NNs) [8,11,54], hidden Markov models (HMM) [55], dynamic-motion-primitive (DMP) models [45], and Gaussian mixture models [14,56]. Concretely, the idea behind LbD is to extract an adequate control law from demonstrations of human motion during ADL tasks.…”

Section: Approaches Based On Learning By Demonstration (Lbd)mentioning

confidence: 99%

“…In the literature, many planning methods based on LbD were proposed for redundant and nonredundant exoskeleton robots [11,14,45,[54][55][56][57]. These methods usually utilize machine-learning methods (supervised and reinforcement learning [58]) to learn the IK of exoskeletons robots.…”

Section: Approaches Based On Learning By Demonstration (Lbd)mentioning

confidence: 99%

“…These DMP parameters are then processed to provide the exoskeleton with reference joint trajectories. In [54], three trained artificial NNs (ANNs) were combined in the form of a closed-loop model that generated elbow angles similar to experimentally recorded trajectories. Garrido et al adopted a planning approach based on LbD [55].…”

Section: Approaches Based On Learning By Demonstration (Lbd)mentioning

confidence: 99%

See 4 more Smart Citations

Motion Planning of Upper-Limb Exoskeleton Robots: A Review

2020

View full text Add to dashboard Cite

(1) Background: Motion planning is an important part of exoskeleton control that improves the wearer’s safety and comfort. However, its usage introduces the problem of trajectory planning. The objective of trajectory planning is to generate the reference input for the motion-control system. This review explores the methods of trajectory planning for exoskeleton control. In order to reduce the number of surveyed papers, this review focuses on the upper limbs, which require refined three-dimensional motion planning. (2) Methods: A systematic search covering the last 20 years was conducted in Ei Compendex, Inspect-IET, Web of Science, PubMed, ProQuest, and Science-Direct. The search strategy was to use and combine terms “trajectory planning”, “upper limb”, and ”exoskeleton” as high-level keywords. “Trajectory planning” and “motion planning” were also combined with the following keywords: “rehabilitation”, “humanlike motion“, “upper extremity“, “inverse kinematic“, and “learning machine “. (3) Results: A total of 67 relevant papers were discovered. Results were then classified into two main categories of methods to plan trajectory: (i) Approaches based on Cartesian motion planning, and inverse kinematics using polynomial-interpolation or optimization-based methods such as minimum-jerk, minimum-torque-change, and inertia-like models; and (ii) approaches based on “learning by demonstration” using machine-learning techniques such as supervised learning based on neural networks, and learning methods based on hidden Markov models, Gaussian mixture models, and dynamic motion primitives. (4) Conclusions: Various methods have been proposed to plan the trajectories for upper-limb exoskeleton robots, but most of them plan the trajectory offline. The review approach is general and could be extended to lower limbs. Trajectory planning has the advantage of extending the applicability of therapy robots to home usage (assistive exoskeletons); it also makes it possible to mitigate the shortages of medical caregivers and therapists, and therapy costs. In this paper, we also discuss challenges associated with trajectory planning: kinematic redundancy and incompatibility, and the trajectory-optimization problem. Commonly, methods based on the computation of swivel angles and other methods rely on the relationship (e.g., coordinated or synergistic) between the degrees of freedom used to resolve kinematic redundancy for exoskeletons. Moreover, two general solutions, namely, the self-tracing configuration of the joint axis and the alignment-free configuration of the joint axis, which add the appropriate number of extra degrees of freedom to the mechanism, were employed to improve the kinematic incompatibility between human and exoskeleton. Future work will focus on online trajectory planning and optimal control. This will be done because very few online methods were found in the scope of this study.

show abstract