AMARSi NAO自适应控制模块集成与原型架构探讨

需积分: 10 95 浏览量更新于2024-07-23 收藏 1.47MB PDF 举报

本篇文档是关于"NAO自适应控制模块架构"的研究报告，由欧盟第七框架计划（EUFP7）资助的AMARSi项目（Adaptive Modular Architectures for Rich Motor Skills，ICT-248311）的一部分，D6.1技术报告，发表于2011年10月，历时18个月。这份报告由赫伯特·雅各布斯大学的赫伯特·雅各布斯、比勒费尔德大学的乔赫恩·斯泰尔和菲利克斯·莱因哈特撰写，旨在概述AMARSi项目中的架构贡献，并在与先验概念架构的视角下进行讨论。报告首先回顾了项目内的现有架构，重点关注那些在早期交付物中有详细记录的部分，强调了它们的简要概述。同时，报告也提及了来自合作伙伴的独立或非AMARSi相关的相关工作，这些工作对理解整个领域的架构发展起到了补充作用。核心目标之一是在认知水平上集成大量模块，实现反应迅速且复杂的（pre-cognitive）行为。M.6.1部分详细阐述了开发出的原型架构及其相对优缺点，通过D6.1交付物的方式进行了验证。报告还着重探讨了架构设计的基本问题，为丰富运动技能架构提出了一种领域特定的语言。通过重新表述AMARSi的典型架构，这份报告尝试提供一个通用框架，以便理解和评估不同模块如何协同工作，以及如何根据实际需求进行适应和优化。此外，报告的公开性（Dissemination level: Public）表明其研究成果具有广泛的传播价值，可为NAO机器人系统的设计者、研究人员和开发者提供宝贵的参考材料。总结来说，这份报告不仅提供了NAO自适应控制模块架构的深入分析，还为设计和实施自适应运动技能的模块化系统提供了关键理论和实践指导，对于理解和改进人机交互、机器人行为控制具有重要意义。

Figure 6: Embedding the SEDS learning module in a basic control architec-

ture (taken from [5]).

which one assumes that motions are driven by an autonomous dynamical

system. In other words, motions are here assumed to follow an ODE that

is time-invariant. Time-invariance is advantageous in that it no longer re-

quires heuristics for time-rescaling when the motion is perturbed. However,

time-invariance in non-linear dynamical systems induces issues of stability

that time-dependent techniques avoid. Methods are developed for learning

an autonomous DS that are asymptotically stable at an attractor. These are

applied to control point-reaching movements. Learning proceeds from a few

demonstration trajectories, by approximating the (assumed ground truth)

ODE via mixtures of Gaussians [6, 7, 8], under strict constraints to ensure

stability, as detailed in Chapter 3 of Deliverable 4.1. The autonomous DS is

embedded into the architecture shown in Figure 6 which provides the infras-

tructure for le arning (supply of training data) and robot control (kinematic

transformations and proprioceptive feedback). The architectures displayed

in Figure 2 and Figure 6 share a common layering of control levels and both

include a kinesthetic, human-robot interaction phase.

Two diﬀerent learning methods have been employed and analyzed:

• Learning the model parameters (means and covariance matrices of par-

ticipating Gaussians) by a constrained nonlinear optimization method

termed SEDS (stable estimator of dynamical systems) leads to a model

with analytically assured global stability (a unique ﬁxed point attractor

at the reaching target location) [7]

• Learning the model parameters via EM, achieving stability in a sub-

region of the total conﬁguration space by an iterated test-reﬁne strat-

egy [6]

A combination of the above methods w ith obstacle avoidance was recently

developed [8].

Main properties of these models are:

• the time-invariant ODE nature of the model directly leads to robust

coping with spatial and temporal disruptions at execution time;

• the methods enjoy assured stability properties;

• the computational (learning) complexity are similar as or better than

other approaches in the ﬁeld.

1.5 Jacobs University

At JAC there are two trains of thought related to architecture. The ﬁrst,

which one might dub ”complete behavior control architectures”, is concerned

with design principles for comprehensive architectures which cover the en-

tire ”cognitive span” between low-level motor control and high-level action

selection and world representation. The second, which one might call ”mod-

ulatable modules”, is concerned with principles of how a complex adaptive

module, which is realized as a recurrent neural network (RNN), can be trained

to become modulatable (i.e. parametrized in a principled way) from higher

levels. Here both lines of research are brieﬂy sketched out.

Complete architectures. This topic has been investigated by H. Jaeger since

about 15 years. The earliest instance is the Dual Dynamics [9] architecture

blueprint for behavior-based robots. Originally conceived in the mid 90’ies, it

has been implemented on a variety of robots in several groups in the context

of RoboCup (mid-size and small-size league), and is still being used at some

places in updated formats. Its practical usefulness was witnessed by several

1rst, 2nd, or 3rd placements in World and European RoboCup championships

by Raul Roja’s group from the TU Berlin [10], and the RoboCup team from

the Fraunhofer Institute AIS. The main characteristics of the Dual Dynamics

designs are the following (compare with Figure 7):

• The main (and single) building block is a behavior. Behaviors are ar-

ranged in layers. Only b ehaviors on the lowest level have access to the

actuators. Higher-level behaviors ”use” lower-level behaviors in the

spirit of ”virtual complex actuators” and control them through select-

ing and activating them in a graded fashion.

剩余60页未读，继续阅读

xyxoswald

粉丝: 0
资源: 1

AMARSi NAO自适应控制模块集成与原型架构探讨

基于LabVIEW的NAO机器人控制研究.pdf

NAO机器人手臂的运动建模与控制.docx

对于NAO机器人控制器

NAO_remote_control:通过遥控器控制NAO

python实现nao机器人手臂动作控制

python程序控制NAO机器人行走

nao_python.zip_NAO迷宫_Nao sonar_nao_nao robot sonar_nao 机器人

nao机器人抓取程序.rar_NAO 抓取_NAO机器人_dayuds_nao robot_nao抓取

python控制nao机器人身体动作实例详解

UBR-1-App:Nao Robot 的 Android 控制器

最新资源